Trial 0010 — 0.3 (map, composable scaffold, recipes), two runs per arm

Question: with hozu map, hozu add feature --with … and the recipes in changing.md, what does the agent cost against Nuxt under the setup of trial 0009?

Targets stated in ADR 0028 before the run (means of two runs):

build ≤ 1.5× Nuxt;
change ≤ 1.5×;
correctness in every run.

Setup

As in trial 0009:
- claude -p (Claude Code 2.1.283) in the app directory, with --setting-sources project,local;
- --model claude-opus-5-5;
- the same prompts (each run gets its own port), the same task, change and hidden acceptance;
- weighted tokens computed from each session's final usage.
Hozu: apps created by the 0.3.0 create-hozu --agent claude, from packed tarballs outside the repository.
Nuxt: the trial 0006 blank scaffold.
Two independent runs per arm and step, all four in parallel. Each change run starts from its own build.
Record: run 1's final apps are in bench/trial/app-0010 and bench/trial/nuxt-0010.

Results

Build:

	Hozu run 1	Hozu run 2	Nuxt run 1	Nuxt run 2
Acceptance	12 / 12	12 / 12	12 / 12	12 / 12
Turns	14	12	5	7
Output tokens	5,169	5,093	5,392	6,209
Fresh input	34,446	33,756	16,177	16,616
Cached input	0.35 M	0.31 M	0.09 M	0.13 M
Weighted	95.2 k	90.2 k	52.1 k	60.9 k
Cost (USD)	0.45	0.43	0.26	0.28

Change:

	Hozu run 1	Hozu run 2	Nuxt run 1	Nuxt run 2
Acceptance	6 / 6, regression 12 / 12	6 / 6, 12 / 12	6 / 6, 12 / 12	6 / 6, 12 / 12
Turns	7	6	6	7
Output tokens	6,745	4,502	4,195	4,333
Fresh input	29,712	24,082	16,194	16,333
Cached input	0.19 M	0.15 M	0.11 M	0.14 M
Weighted	82.8 k	61.2 k	48.3 k	51.6 k
Cost (USD)	0.41	0.31	0.24	0.25

Against the targets (means):

	Target	0.2 (trial 0009)	0.3 (this trial)	Met?
Build	≤ 1.5×	2.05×	1.64× (92.7 k vs 56.5 k)	no
Change	≤ 1.5×	2.34×	1.44× (72.0 k vs 49.9 k)	yes
Correctness	every run	yes	yes, 8 of 8 runs	yes

App source: Hozu run 1 is 16.4 KB, Nuxt run 1 is 9.3 KB.

Where the tokens went

Hozu build (both runs):

the skill, then the app's files and patterns.md (11.9 k characters);
then hozu add feature tasks --page / --with detail,toggle,filter,remove, followed by printing every generated file (16.3–16.5 k);
then small edits to the spec's texts, and hozu check;
then hozu get / hozu post.

Neither run rewrote the scaffold, which is why output tokens halved against trial 0009 (10.3 k → 5.1 k). Both runs still started the server once, to check attributes such as aria-pressed that hozu get does not print.

Hozu change (both runs):

changing.md with the app's files (17.1 k);
then the recipe section and the app's views (12.8–14.1 k);
then one edit, hozu check and a post flow.

Run 1 also reread part of a file (7.2 k).

The change went from 15 turns in trial 0009 to 6–7 here, and the server was never started.

Reading the result

The change now costs about what the targets asked for. The recipes turned the change into one edit, and --button and map removed the detours of trial 0009.
The build is close but over. What remains is reading:
- the scaffold's output is printed in full to be edited (16 k);
- the app's own files and patterns.md are read before scaffolding (12 k).
The model writes Nuxt with 3 k of reading.
Correctness stayed equal in all eight runs.

Conclusion

Measured: build 1.64× (target missed), change 1.44× (target met), correctness equal. From 0.2 to 0.3, build fell from 2.05× to 1.64× and change from 2.34× to 1.44×.
Candidates for the next step (not decided):
- hozu get --attrs (or a selector), so attributes can be checked without a server;
- a scaffold that prints a short summary of what it generated instead of prompting a full read;
- tell agents to skip patterns.md when a scaffold part covers the pattern.