description: "Probe the review agent's credentials end to end, then review the repository's code."
params:
FOCUS:
type: string
default: "correctness bugs, security problems, and resource leaks"
description: "What the review should concentrate on"
steps:
probe_credentials:
agent:
harness: claude
model: sonnet
prompt: |
This is a credentials probe, nothing more. Confirm you are executing inside the
container by writing the JSON object {"ok": true} to the file named by the
SCSH_RESULT environment variable, then stop. Do nothing else — no commits, no
network, no other files.
output:
ok:
type: bool
review:
needs: probe_credentials
when:
probe_credentials.ok: "true"
agent:
harness: claude
model: sonnet
inputs:
FOCUS: params.FOCUS
prompt: |
Review the code in this repository, concentrating on the FOCUS input. Read the
most important source files first (entry points, then whatever they pull in),
then write a JSON object with two string fields to the file named by the
SCSH_RESULT environment variable: "summary" (a few sentences on the overall
state of the code) and "findings" (a numbered list in one string, most severe
first; the string "none" when you found nothing worth raising). Make no commits.
output:
summary:
type: string
findings:
type: string