Trial 0016 — 0.5 with the trial 0015 fixes, four runs per step, and a second agent

Question: trial 0015 measured the 0.5 build at 1.46× Nuxt, but its change step was noise (one outlier at 177 k), and it found four sources of friction. After fixing them:

&@Button in hozu post;
comma lists in --select;
a message for a redirected post;
patterns topics that carry their own snippets;

what does 0.5 cost with four runs per step, and does a second agent (Codex) get it right?

Setup

Identical to trials 0014 and 0015:
- the notes spec, change request, prompts, model (claude-opus-5-5) and hidden acceptance;
- Nuxt is compared with trial 0012's two runs: build 74.8 k, change 61.1 k. Nuxt was not re-run.
Hozu: four Claude runs on the packed 0.5.0 (19 packages).
Codex: one run with codex exec -s workspace-write, in an app created with --agent agents (it reads AGENTS.md), and the same prompts.

Results

Correctness: every run passes everything.

	h1	h2	h3	h4	Codex
Build (15)	15	15	15	15	15
Change: new (6)	6	6	6	6	6
Change: regression (15)	15	15	15	15	15

Cost (Claude, weighted tokens):

	h1	h2	h3	h4	Mean	vs Nuxt	Trial 0014	Trial 0015
Build	87.5 k	103.0 k	101.0 k	122.1 k	103.4 k	1.38×	1.80×	1.46×
Change	74.5 k	101.4 k	82.7 k	96.7 k	88.8 k	1.45×	1.81×	2.36×
Build + change					192.2 k	1.41×	1.80×	1.86×
Calls, build / change	12 / 11	20 / 13	13 / 10	19 / 12	16.0 / 11.5	6 / 8.5	18 / 12	16 / 17

Codex:

It used 14 commands to build and 14 to change: 638 k and 626 k input tokens, of which 591 k and 579 k were cached, and 4.8 k and 4.6 k output tokens.
Its model and pricing differ, so it is not compared with Nuxt. The point is that a second agent, reading AGENTS.md, built and changed the app with every check passing.

Anatomy (bench/trial/anatomy.mjs, means, trial 0014 → 0016):

	Total	Calls	Output × 5	Docs, with carry	Docs calls	Read app	Verify calls
Build	134.7 → 103.4 k	17.0 → 12.5	36.5 → 28.2 k	29.9 → 28.3 k	5.5 → 3.3	14.5 → 8.4 k	3.0 → 3.3
Change	110.6 → 88.8 k	12.0 → 11.0	36.3 → 28.4 k	32.5 → 22.6 k	4.0 → 3.0	4.8 → 4.2 k	3.5 → 2.5

Reading the result

Correctness is unchanged at the top.
- That is 180/180 checks across five runs and two agents: isolation, double submit, no-JS forms, HttpOnly, pinning, in-browser search, and every regression check.
- Trial 0012's Nuxt run 1 broke three features on the same change.
Cost fell on both steps against trial 0014 (0.5 before R1–R3):
- build −23 %, change −20 %, together −22 %;
- build + change is now 1.41× Nuxt, the lowest on this task (trials 0012 2.46×, 0014 1.80×, 0015 1.86×).
Where it fell:
- fewer calls (a build is 12.5 calls against 17);
- less output, because assignments and operators are shorter than op.*;
- fewer lookups in the guide;
- less reading of the scaffold.
The docs cost per run is still the largest single part (22–28 k). Carried through every call, it is what keeps Hozu above Nuxt, whose agents read nothing before writing.
Four runs per step still vary by about ±15 %. Trial 0015's change outlier (177 k) came from its own failed scripted edit and from the CLI friction fixed here; no run in this trial went above 122 k.
Every run still starts a server once at the end. The prompt offers a port for it, and Nuxt runs do the same.

Conclusion

0.5 (R1–R3 plus the trial 0015 fixes) costs 1.38× Nuxt to build and 1.45× to change, with every check passing for two different agents.
The ADR 0038 target (≤ 1.3×) is not met, but the remaining gap is the guide itself.
Measured against 0.4-era Hozu on the same task, an agent now spends about a fifth less.