Real jobs, recorded from the request to the merge

Each case is one run of production Coroid on a real repository, written up as it happened: the request, what QA and review found, what a person sent back, and what it cost. Nothing was run again to hide a failure.

Most cases run on Coroid Expo, a 98 000-line product built with Claude Code and Codex. Its repository is private, so its pull requests are named, not linked.

Close a security gap

Close a hardening gap from the threat model
Coroid Expo's content security policy still allowed inline styles, a relaxation its own threat model listed. Coroid removed it without breaking the fonts visitors read in Arabic, Japanese and Chinese, and QA proved it in a real browser.
To the pull request
43 min
Criteria verified
8 of 8
Model spend
0,51 $
Fail closed when a dependency is down
When Coroid Expo's moderation model was unavailable, feature requests and questions to its persona went through with only a word filter in front of them. Coroid made both fail closed, with tests, and Quality Control sent the first round back before any pull request existed.
To the pull request
1 h 23 min
Criteria verified
10 of 10
Model spend
1,80 $

Pay down test debt

Pay down test debt
Three end-to-end journeys in Coroid Expo had been switched off, red by design until their features landed. Coroid turned on the two that can pass, left the third off and named exactly what it still waits for, without weakening an assertion.
To the pull request
57 min
Model spend
0,50 $

Keep the supply chain safe

Pin CI actions and keep them current
Coroid Expo's workflows referenced GitHub Actions by tag, and its deploy job runs with registry and cluster credentials. Coroid pinned every action to the commit of the release already in use and added grouped weekly updates.
To the pull request
26 min
Criteria verified
6 of 6
Model spend
0,15 $

Make it accessible

An accessibility pass, as a phased plan
One request asked for an accessibility audit of Coroid Expo beyond what axe checks, the fixes, an end-to-end test for each criterion and an accessibility statement. Coroid planned it in phases, ran the tasks, and verified every requested criterion twice: in each task, and again on the combined result.
To the pull request
23 h 59 min
Criteria verified
8 of 8
Model spend
14,42 $

Work next to your developers' agents

A Claude Code pull request meets the team's quality bar
A developer fixed a voice-token authorization gap with Claude Code and opened the pull request themselves. Coroid checked it with the same Quality Control it applies to its own work, and Coroid Review read it. From the same session, the developer asked Coroid over MCP for tests, and Coroid delivered that pull request too.
To the pull request
31 min
Criteria verified
8 of 8
Model spend
0,26 $

Ship a release

Ship a release
Coroid Expo had last been deployed by hand, and Coroid had no record of what ran in production. Coroid shipped the merged work as one release: checks, a person's approval, a deploy through Expo's own workflow, the run verified on GitHub, and health judged after it.
To the pull request
39 min
Model spend
0,28 $

Onboard a repository

Onboard an existing repository
We signed up as a new organization on the Free plan, with a private copy of Coroid Expo in its own GitHub organization. Claude Code followed Coroid's setup runbook, a person did only the steps a person must, and Coroid's first task ended in a pull request.
To the pull request
25 min
Criteria verified
5 of 5
Model spend
0,00 $

Build a product from scratch

Chrono City: от задания до проверенных pull-request'ов в pull request
Сцена города из пяти эпох казалась готовой после первой сборки. При проверке в браузере были выявлены два невыполненных критерия; работа вернулась на доработку, после чего pull request был готов. Это отчёт об этом процессе.
To the pull request
1 h 48 min
Criteria verified
11 of 11
Model spend
1,00 $