pretty_name: Drupal module pick, with and without a module docs server
license: other
license_name: gpl-2.0-or-later
license_link: https://www.gnu.org/licenses/old-licenses/gpl-2.0.html
tags:
- evaluation
- agents
- drupal
- mcp
configs:
- config_name: default
data_files:
- split: full
path: data/full.jsonl
Drupal module pick
Which module does a coding agent pick when a Drupal ask names none, and does a
module documentation server change the pick? Each run gives Claude Code
(2.1.303, model claude-sonnet-5-5) one outcome-only task on a
real Drupal CMS site, then reads what it changed: Composer packages, enabled
modules, config and custom code. The tasks, harness and grader are in
evals/module-pick/ of drupal/ai_best_practices (https://git.drupalcode.org/project/ai_best_practices/-/work_items/3589009).
| A0 (no docs server) | A1 (docs server connected, the prompt does not mention it) | A2 (docs server connected, the prompt names it) | |
|---|---|---|---|
| Good pick, strict | 37 of 60 | 45 of 60 | 44 of 60 |
| Good pick, lenient | 44 of 60 | 50 of 60 | 50 of 60 |
| Custom code only | 13 | 8 | 2 |
| Off-list contrib | 1 | 2 | 6 |
| Runs that used the server | 0 | 46 | 48 |
A1 against A0, paired by task: strict, 10 tasks good only with the server and 2 only without it (exact McNemar p = 0.039); lenient, 8 against 2 (p = 0.109).
"Strict" counts runs that required, enabled or configured a module from the task's good list. "Lenient" also counts runs where the bed already had a good module and the answer names it. The good lists are a reading of the module ecosystem, not a standard.
Limits: one model, 1 run per task and arm, one bed.
| Path | What |
|---|---|
data/full.jsonl |
One row per run: task, arm, outcome, what changed, server calls, cost |
runs/<run>/ |
Transcript (stream-json), grade, and site snapshots before and after |
eee/module-pick/<developer>/<model>/ |
Every Eval Ever 0.2.2 record, strict and lenient rate per arm |