Trial 0015 — The notes app on 0.5 with ordinary TypeScript and an indexed guide (ADR 0038, 0039)

Question: trial 0014 left the cost at 1.80× Nuxt to build and 1.81× to change, and ADR 0038 traced most of the gap to reading the guide. After R1 (an indexed guide and hozu docs), R2 (fewer calls) and R3 (ordinary TypeScript), does the cost fall, and does correctness hold?

Setup

Identical to trial 0014:
- the same spec, change request, prompts, model (claude-opus-5-5) and acceptance;
- Nuxt is compared with trial 0012's runs: build 74.8 k, change 61.1 k.
Hozu: two runs on the packed 0.5.0, 19 packages including @hozu/transform.
The Codex run is void again: it hit its usage limit ("try again at 3:55 AM") before its first command.

Results

Correctness:

	Hozu run 1	Hozu run 2
Build (15 checks)	15	15
Change: new behaviour (6)	6	6
Change: regression (15)	15	15

Cost (weighted tokens):

	Run 1	Run 2	Mean	vs Nuxt	Trial 0014
Build	122.4 k	95.8 k	109.1 k	1.46×	134.7 k (1.80×)
Build calls	17	15			21 / 15
Change	177.3 k	111.2 k	144.2 k	2.36×	110.6 k (1.81×)
Change calls	21	13			13 / 11
Build + change	299.7 k	207.0 k	253.3 k	1.87×	245.3 k (1.81×)

Build anatomy (bench/trial/anatomy.mjs, means, trial 0014 → 0015):

	Total	Calls	Output × 5	Docs, with carry	Docs calls	Edit calls
0014	134.7 k	17.0	36.5 k	29.9 k	5.5	4.5
0015	109.1 k	14.0	29.5 k	22.2 k	3.0	2.5

Where the tokens went

Build:
- both runs loaded the 4.6 k skill, ran hozu docs auth (and contracts or forms), then scaffolded;
- they then read every generated file (13–16 k), edited, verified with hozu post, and finished with one server and curl check;
- docs lookups fell from 5.5 calls to 3, and writing fell by a fifth.
Change, run 1 (177 k, 21 calls):
- a scripted edit failed halfway (an assertion in its own Python), and repairing it took several calls;
- hozu post then cost four more calls:
  - --next 'POST / id=n2&@Pin' failed with Field "" must be name=value. This is a bug: run 2 of the build hit it too;
  - --select 'input,label' is unsupported;
  - post / while signed out said only "has no form that posts".
Both change runs read the bookmarks example (about 12 k), because topics/patterns.md points to it for the patterns it marks (example).

Reading the result

Correctness held: 72/72 checks again, as in trials 0012–0014.
The build fell by 19 %, in the components R1–R3 target:
- fewer guide lookups and less carried documentation;
- fewer edit calls, and a fifth less output.
- The ADR 0038 goal of ≤ 1.3× is not met, but 1.46× is the lowest build ratio on this task (trial 0012: 2.79×, trial 0013: 1.66×, trial 0014: 1.80×).
The change did not fall.
- Run 2 (111 k) matches trial 0014.
- Run 1 is a 177 k outlier, caused by its own failed edit and by CLI friction that the guide sent it into.
- With two runs, the change result is noise around 1.8–2.4×, not a measured regression or improvement.
Build + change together are equal to trial 0014 (253 k against 245 k). 0.5 has not yet shown a net saving over the whole task.

What this trial found to fix

hozu post --next 'POST / a=1&@Button' must accept a button after &.
--select should accept a comma-separated list.
post to a page that redirected to sign-in should say so.
topics/patterns.md should carry each pattern's snippet itself instead of pointing to the example files.