Managed Code Review gives every review agent this file's contents as the
repository's review instructions. Its citations, the
standards/review/<file>.md#<heading> paths in parentheses, are not include
lines, so Managed Code Review does not read the cited files and sees only the
words printed below (see "Depth"). A self-hosted or local review running the
review plugin also follows each citation into the standards repository
for the full criterion behind the line.
Review in this organization splits into two mutually exclusive scopes, so a
finding belongs to exactly one and is never reported twice. The code-review lane
(pr-review, and Managed Code Review) reads this file and applies the
code-review scope below. The security lane (pr-review-security) runs
from its own security-only prompt, not this file; the security scope below
records the split for every surface that does read this file, so the
code-review lane knows what to leave to the security lane, and a self-hosted
or local review (for example a review-plugin security agent) applies the
right section.
This organization's criteria (standards/review/ in standards) resolve
to three severities: Critical, Important, Suggestion. On this
surface's three markers:
| SSOT severity | Managed marker |
|---|---|
| Critical | 🔴 Important |
| Important | 🔴 Important |
| Suggestion | 🟡 Nit |
🟣 Pre-existing is not an SSOT severity: it is this surface's own detection-time judgment of whether a finding predates the diff, orthogonal to the three severities above, and no criterion assigns it.
A finding tagged blocking is always 🔴 Important here, whatever severity its
criterion would otherwise carry in context. blocking is this file's own
tag, not a severity a standards/review/ criterion assigns.
Managed Code Review does not follow a citation below: a citation is not an
include line
(Include other files in REVIEW.md),
so the cited file never reaches the review. A self-hosted or local review
using the review plugin follows one into the full criterion, its severity
nuance, and its sources, resolving the path from the standards checkout
root: this repository's own root when standards reviews itself, or the
--add-dir mount root everywhere else.
A reviewed repository's docs/review/README.md, when present, indexes its
own per-concern review criteria, and a self-hosted or local review also
reads the files there whose concern the change touches. Those paths resolve against
the reviewed repository's root, not the standards checkout or its
--add-dir mount.
This lane owns every review dimension except security: correctness,
design, conventions, error handling, observability, tests, and
documentation. On a repository whose CI runs the security lane (a
.github/workflows/pr-review-security*.yml workflow exists), it does
not report security findings, such as vulnerabilities, authorization
or tenancy gaps, credential exposure, or injection: each belongs to that
lane alone and is omitted here even when a hunk plainly contains one.
On a repository with no such workflow, no other lane would report them:
report security findings here too, applying the security-scope checks below.
Always check:
- A high-risk security action, such as authentication, an authorization
failure, a privilege change, or sensitive-data access, has a matching
audit-log entry (
standards/review/observability.md#logging). This checks logging completeness; whether the action itself is safe is the security lane's question. - A change that writes two or more related records, files, or state
locations carries an atomicity mechanism spanning them: a transaction,
an atomic rename, a constraint, or a compensation step. An
interruption between steps must not leave state no code path expects
(
standards/review/error-handling.md#atomicity). - No Dependabot pull request waits on a human. One whose updates are all
non-major, with GitHub Actions only from
actions/*,github/*,anthropics/*or the organization's ownmelodic-software/*, and none in the policy'sreviewOnlysecurity-critical list, merges onceCI Statuspasses. Any other, a major included, merges on an AI review lane'smergeverdict at its current head onceCI Statuspasses, and a verdict on a security-critical dependency cites the threat-model control it checked. A change that adds an action from another publisher says why no vendor action orghor script step serves (components/dependabot-policy/README.md#merge-tiers-for-dependabot-pull-requests).
This lane reports only security findings: vulnerabilities, missing authorization or tenant scoping, exposed secrets or credentials, injection, and their direct enablers. Everything else, including style, design, correctness with no security impact, and test coverage, belongs to the code-review lane; omit it.
Always check:
- A handler or endpoint that receives an object id (a path segment, body
field, or query parameter) checks the caller is authorized for that
specific record, not only that the id is well-formed. Tagged
blocking(standards/review/security.md#object-level-authorization). - A data-access path that reaches tenant-owned data carries an explicit
tenant scope. Tagged
blocking(standards/review/multi-tenancy.md#tenant-scoped-data-access). - No secret, token, or credential is added to tracked source or config.
Tagged
blocking(standards/review/security.md#secrets-and-credentials). - A query or process invocation built from external input is parameterized,
never string-concatenated. Tagged
blocking(standards/review/security.md#trust-boundaries-and-injection).