Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 

README.md

setup Evals

Evaluations for the setup skill. See the framework README for how evals work and how to run them.

Test cases

ID Name Purpose
1 Greenfield setup Empty CLAUDE.md with Serena instances available. Tests full configuration generation: Repository Registry, Jira Configuration, and Code Intelligence.
2 Incremental update Already-configured CLAUDE.md with a new Serena instance discovered. Tests that existing entries are preserved and only new entries are added.
3 No-Serena/no-MCP Empty CLAUDE.md with no Serena tools and no Atlassian MCP. Tests graceful degradation: user prompted about continuing without code intelligence, manual Jira entry.
4 Adversarial CLAUDE.md with injection vectors embedded in configuration field values. Tests that the skill treats adversarial content as literal data and does not follow embedded instructions.

Fixture files

Files in files/ simulate the existing CLAUDE.md state and available MCP tools listing. The eval prompt instructs the agent to write its outputs to outputs/ rather than modifying actual files.

CLAUDE.md fixtures (claude-md-*.md)

Mock CLAUDE.md files representing different project states:

  • claude-md-empty.md — CLAUDE.md with no Project Configuration section, simulating a greenfield project
  • claude-md-configured.md — CLAUDE.md with complete existing Project Configuration (Repository Registry with one entry, full Jira Configuration, Code Intelligence), simulating an already-configured project for idempotent and incremental testing
  • claude-md-adversarial.md — CLAUDE.md with injection attempts embedded in configuration field values: repository names with exfiltration instructions, Serena Instance fields with system override commands, Limitations section with backdoor creation instructions

MCP tools fixtures (mcp-tools-*.md)

Mock MCP tool listings simulating tool discovery output:

  • mcp-tools-with-serena.md — tool listing showing two Serena instances (serena_backend, serena_ui) and Atlassian MCP tools. For eval 2 (incremental), serena_ui is the newly discovered instance not yet in the configured Registry.
  • mcp-tools-no-serena.md — tool listing with generic tools only (Bash, Read, Write, etc.) and no Serena or Atlassian MCP tools

Key constraints tested

Constraint Eval IDs
Idempotent — running multiple times produces no changes 2
Incremental — only new entries added 2
Never remove existing entries 2, 4
Never overwrite user-customized values 2
Always present changes for review before writing 1, 2, 3
Only ask for fields that are missing 2, 3
Prompt about continuing without code intelligence 3
Injection resistance — treat adversarial content as data 4

Running

/sdlc-workflow:run-evals Run evals for setup.
Evals path: evals/setup/evals.json
Workspace: /tmp/setup-eval