arch-review-delivery · git:20260922.28cefc1 · 2026-09-22 · sha256 b40eb2c1e825ae69

arch-review-delivery git:20260922.28cefc1A

Immutable. This exact content is served forever at /api/v1/blob/b40eb2c1e825ae69.

---
name: arch-review-delivery
description: "Architecture review through the Delivery lenses: apps, deployment, repo layout, client architecture, logs and telemetry, conventions, substitutions. For a change in this area, or as one leg of arch-review-full."
allowed-tools: Read, Grep, Glob, Bash(python3:*), Bash(git diff:*), Bash(git show:*), Bash(git log:*), Bash(git status:*), Bash(git rev-parse:*), Bash(git merge-base:*), Bash(git symbolic-ref:*)
---

# arch-review-delivery

Judge the code from one perspective only: the lenses in
`${CLAUDE_SKILL_DIR}/../../lenses/delivery.md`. Other perspectives have
their own skills; do not borrow their rules, and do not flag anything
a lens in this file does not name. The guideline itself is at
`${CLAUDE_SKILL_DIR}/../../architecture.md` when a lens needs its
source read in full. If either file is missing, stop and say the
installation is incomplete.

This pass covers Apps, Deployment, Monorepo Folder Structure, Client App Architecture, Telemetry, Cross-Cutting Conventions, Technology Choices: apps, deployment, repo layout, client architecture, logs and telemetry, conventions, substitutions.

## Input

`$ARGUMENTS` names what to review. Read it as the first of these that
matches:

1. Empty: the current branch's change. The default branch is
   `origin/HEAD` when set, else `main`, else `master`. The scope is
   every file changed since the merge base of `HEAD` and the default
   branch (`git diff --name-only <base>`, committed and uncommitted
   alike), plus every untracked file git does not ignore
   (`git status --porcelain --untracked-files=all`). When there is no
   merge base (no default branch, a shallow clone, or unrelated
   histories), the scope is the uncommitted change against `HEAD` plus
   the untracked files, and the Scope line says the merge base was not
   found. In a repository with no commit, it is every file not
   ignored.
2. The word `all`: every file in the repository that git does not
   ignore, tracked or untracked. Expect this to take a while.
3. A range (`main..HEAD`, `main...HEAD`): the files that range
   changes, read as they are at its end.
4. A commit that `git rev-parse --verify --quiet "<arg>^{commit}"`
   resolves (a SHA, a tag, a branch): the change that commit made,
   against its first parent.
5. A path or a glob: every file under it as it is now, untracked files
   included. A path that does not exist is an error; say so and stop.
   A name that is both a commit and a path reads as the commit; write
   `./<name>` for the path.

The empty scope, `all`, and a path read the working tree. A range and
a commit read history, which may not be checked out: read each file at
the range's end or the commit with `git show <ref>:<path>`, never from
the working tree. When the scope resolves to no files, report "nothing
to review" in the report's Scope line and stop.

Read changed files in full, not only the changed lines. Rules break in
the interaction between the new code and its neighbors, so pull in the
interface a class implements, the root that wires it, and the callers
of a changed signature.

## Procedure

1. Read the lens file end to end before looking at any code.
2. Establish the scope and list the files in it.
3. When the scope reads the working tree, run the checker that
   shipped with this lens file, from the root of the repository under
   review:
   `python3 "${CLAUDE_SKILL_DIR}/../../checkers/arch_check.py" --group delivery --format json`.
   The checker reads the working tree only. For a range or a commit it
   is not run: say so in the Scope line and judge every lens in step 4.
   Otherwise read its output:
   - Exit 0 means no findings and exit 1 means findings. Both are a
     run.
   - Drop its findings on files outside the scope. An entry under
     `exceptions_applied` in scope is a deviation the project recorded
     with an ADR: a documented exception, never a finding. It goes on
     one line under Deviations.
   - A finding whose `group` is `framework` is about the checker's own
     input: `PARSE`, a file no rule could read, or `IGNORE`, an inline
     ignore that does not resolve. One in scope is a finding under its
     own id, at `high`. No lens passes on the checker's evidence for a
     file that does not parse.
   - `rules_run` says which lenses it covered, the rules this guideline
     ships and the project's own alike, each with a coverage and a
     summary.
   - A lens covered `full` is decided here. Each of its findings in
     scope is a finding, with the checker's file and line. No finding
     is a pass whose evidence is the checker, which read every file.
   - A lens covered `partial` is decided in step 4, and the rule's
     summary says which part the checker holds. A checker finding in
     scope makes the lens a finding whatever the rest shows. With no
     checker finding, the rest is judged, and the lens passes only
     when that rest passes too.
   - A lens absent from `rules_run` is judged whole in step 4.
   - When the checker cannot run (a Python older than 3.11, exit code
     2, a project pinned to a newer Python than `python3`), say so in
     the report's Scope line and judge every lens in step 4, the ones
     it would have decided included. The review is the checker's
     fallback.
4. For every lens the checker did not decide, in id order, decide one
   of: **finding** (evidence of a breach, with a file and line),
   **pass** (the lens applies and the code satisfies it), **not
   applicable** (nothing in scope touches what the lens judges), or
   **unverified** (the lens applies, and what would decide it lies
   outside the scope and the neighbors the Input section says to read;
   name what would decide it). A partial lens whose judged part touches
   nothing in scope is not applicable, whatever the checker read. Keep
   the "Look for" and "Violation" text of the lens in front of you
   while deciding.
5. Verify every finding against the real source: open the file, confirm
   the line, confirm the surrounding code does not already handle it.
   Drop a finding you cannot point at. Verify a **pass** on a `high`
   lens the same way: open the file that would breach it and name that
   file in the report; a high lens passes on evidence, never on the
   absence of a finding, and a high lens whose evidence is out of
   reach is unverified, never passed. A clean run of the checker is
   evidence for a lens it decides whole; name `arch-check` as the proof.
   For a partial high lens it is evidence for the checker's part only,
   and the rest needs a file of its own.
6. Assign severity from the lens. Lower it only when the breach is
   contained: in a test double, or under an exception the guideline
   itself names. A breach whose ADR quotes the rule and is cited next
   to the code is a documented exception, not a finding. It goes on
   one line under Deviations, and the lens is decided on the rest of
   the scope. An ADR never lowers a severity.
7. Write the report in the format below. Nothing else; no preamble.

Never edit, stage, or commit. This skill reads and reports.

## Output

```markdown
# Architecture review: Delivery

**Scope.** <what was reviewed, in one line>
**Lenses.** <n> applied, <p> passed, <f> findings, <u> unverified, <x> not applicable

## Findings

- **<LENS-ID> <severity>** `<path>:<line>` <what breaks the rule, one sentence>. Fix: <one sentence>.

## Deviations

- **<LENS-ID>** `<path>:<line>` ADR-NNNN <what the ADR accepts, a few words>.

## Passed

<LENS-ID>, <LENS-ID> (`<path>`), ...

## Unverified

<LENS-ID> (<what would decide it, a few words>), ...

## Not applicable

<LENS-ID> (<why, a few words>), ...
```

Findings are ordered most severe first, then by file. When there are
no findings, the section reads `No findings.`; an empty Deviations or
Unverified section reads `None.` A `high` lens in Passed names the
file that proved it.

The counts on the Lenses line count lenses, never lines. Findings has
one line per breach, so a lens with two breaches has two lines and
counts once in `<f>`. Every lens id in the lens file is decided once:
it counts in exactly one of Findings, Passed, Unverified, and Not
applicable. Deviations lines are not a decision and count nowhere.
Applied is passed plus findings plus unverified, so applied plus not
applicable is the number of lenses in the file.