Build a Five-File Canary Before You Trust an Agentic Coding Product

I do not need another agent leaderboard. I need a tiny repository I can run through every product before trusting it with a real codebase.

My canary has five files:

AGENTS.md             one explicit constraint
src/price.js          a small bug
test/price.test.js    one failing test
ADR.md                a stale design decision
tools/check.sh        a permitted verification command

The task is deliberately specific: fix rounding without changing the public API, update the stale ADR only if behavior changes, run the checker, and stop before committing.

Score observable evidence, not how confident the chat sounds:

Capability	Evidence
Planning	names the constraint and affected files
Tool use	executes only the documented checker
Correction	uses the failed assertion to revise code
Context	preserves the public API
Human handoff	presents diff and stops before commit

Add one hostile fixture: make the first test failure misleading. The agent should inspect the implementation and constraints rather than patching the assertion until it turns green.

Anthropic’s Building Effective Agents distinguishes workflows with predefined paths from agents that dynamically direct their own process. This canary tests the product behavior either way without depending on its marketing label.

MonkeyCode is a useful candidate because it is open source and has a browser-based SaaS platform currently free to start. I can run the disposable canary without treating free access as permanent or claiming a result before collecting evidence. The same repository stays portable for the next tool.

I am a MonkeyCode user, not affiliated with the project. This account is managed by the same operator as the other accounts in this batch.