Whole-repo architectural drift sweep. Spawns eight parallel read-only subagents, each focused on one axis across the entire codebase (not just the diff). Aggregates findings into a single severity-ranked report. Use periodically, before milestones, or on user /audit-architecture.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Whole-repo architectural drift sweep. Spawns eight parallel read-only subagents, each focused on one axis across the entire codebase (not just the diff). Aggregates findings into a single severity-ranked report. Use periodically, before milestones, or on user /audit-architecture.
audit-architecture
Whole-tree non-deterministic audit. Where /review checks the diff, this checks every file in the repository for slow drift accumulated across many small changes.
When to invoke
User /audit-architecture.
Scheduled drift sweep (weekly).
Before milestones / version bumps.
After large refactors.
How it runs
Determine the repository root from the current working directory: git rev-parse --show-toplevel (or use the cwd if not a git repo). All paths in the prompts are repo-relative — every subagent inherits the cwd.
Spawn eight read-only subagents in parallel in a single message (multiple Agent tool calls in one block). Each is subagent_type=Explore. Each prompt below is self-contained — copy verbatim. Subagents have no conversation context.
After all eight return, build the aggregated report (format at the bottom).
Subagent 1 — Layering, import direction & package boundary
Audit packages/client/src/ and packages/cli/src/ in the current repository for layering violations. Use ripgrep / file reads — read-only, no edits.
Skip *.test.ts files entirely — per CLAUDE.md, "Layering & boundary rules are declared per package and apply to production source only; *.test.ts files are exempt and may import across layers." Audit only production .ts files.
This is a bun workspace: packages/client (@metabase/client) is the private Metabase API client, packages/cli (@metabase/cli) is the publishable CLI built on it.
Package boundary (hardest rule):
Nothing under packages/client/src/** may import from packages/cli — not by an @metabase/cli specifier, and not by a relative path that resolves outside the client's own src tree (../../cli/src/... reads as ordinary navigation but lands in the other package).
The client's dependency budget is zod, semver, and node: builtins. Root devDependencies hoist, so any third-party import resolves at runtime — report every production client file importing a package outside that budget.
packages/cli/src/** consumes the client by specifier: @metabase/client/<path> (e.g. @metabase/client/domain/card, @metabase/client/json, @metabase/client/errors). A relative path from the CLI into packages/client/src/ is a violation.
Within packages/client/src/:
domain/** may import only from zod and other domain/** files.
resources/** — one file per Metabase resource, holding every /api/ path the client sends — may be imported only by client.ts and by sibling resources/** files. A domain/**, http/**, auth/**, version/** or root-level client file importing a resource is a violation; the direction runs one way, and the permitted importers are an allowlist, so a directory added later is not exempt for having arrived after the rule.
Everything else may import freely within the client, subject to the boundary rules above.
Within packages/cli/src/:
commands/** may import from any other CLI location and any @metabase/client/* path. It reaches Metabase only as client.<resource>.<method>(…) — a command naming an /api/… path literal or calling requestParsed / requestRaw / requestStream / paginatePages is a violation. Both rules are absolute and admit no exception, so report every offending file. Report separately any command that reaches the network by another name: commands/auth/login.ts's tryDiscoverMetadata is the one sanctioned case, and no pattern rule can see a second one.
core/** may import from runtime/**, other core/**, and @metabase/client/*. Never from commands/ or output/.
output/** may import from runtime/**, core/** except core/auth/, and any @metabase/client path except @metabase/client/http and anything beneath it (presentation consumes the error taxonomy from @metabase/client/errors, never the HTTP boundary). Never from commands/.
runtime/** may import from @metabase/client/* (typed errors are the project's pattern: runtime/body.ts uses parseJson and ConfigError, runtime/yaml.ts uses ValidationError). It must not import from commands/, core/, or output/.
For every violation, report path:line: <import-stmt> — violates <rule>. If zero violations, output the single line (clean).
Subagent 2 — Type strictness
Audit packages/client/src/, packages/cli/src/, and tests/ in the current repository for type-strictness violations. Read-only.
Patterns to find:
as <Type>, as unknown as, as never, as any (any cast form).
: any or <any> annotations.
Record<string, unknown> used as a typed value (the sole sanctioned home is packages/cli/src/output/projection.ts — flag every other instance and let the aggregator decide).
// @ts-ignore, // @ts-nocheck, // @ts-expect-error (allowed only when paired with // reason: and // FIXME(date)).
For each, report path:line: <pattern>: <snippet>. If zero violations, output (clean).
Subagent 3 — Boundary & process discipline
Audit packages/client/src/ and packages/cli/src/ in the current repository for boundary violations. Read-only. Skip *.test.ts files entirely (production source only — tests are exempt).
The permitted homes are declared per package in CLAUDE.md — a pattern's home in one package says nothing about the other. Judge each file against the table for the package it lives in. All paths below are relative to that package's src/.
The client — every file under packages/client/src/:
Pattern
Allowed only in
bare fetch(, globalThis.fetch, third-party HTTP libs (got/axios/node-fetch/undici)
http/**
new URL(
http/**, url.ts
JSON.parse(
json.ts
setTimeout inside an await wait loop
poll.ts, http/retry.ts
process.stdout.write, process.stderr.write
nowhere — the client owns no output
process.exit(
nowhere — the client never ends the run
@clack/prompts, @clack/core
nowhere — prompting is the CLI's job
fs, fs/promises, node:fs, node:fs/promises
nowhere — the client touches no files
child_process, node:child_process
nowhere — the client spawns nothing
an import specifier ending in .ts
nowhere — imports are extensionless
The CLI — every file under packages/cli/src/:
Subagent 4 — Test parity
Audit pairing between the two packages' source and their tests in the current repository. Read-only.
Test layout: this project uses two tiers — colocated unit tests at packages/<p>/src/**/*.test.ts (beside the code, inside the package that owns it) and end-to-end at the repo root under tests/e2e/<noun>.e2e.test.ts (workspace-wide, driving the built packages/cli/dist/cli.mjs). There are no tests/integration/ or tests/smoke/ directories, and neither package has its own e2e tier.
Required pairings:
For every leaf command at packages/cli/src/commands/<noun>/<verb>.ts: a tests/e2e/<noun>.e2e.test.ts exists and exercises that subcommand via runCli. Adding a new command without an e2e test is a missing pairing.
For every leaf command at packages/cli/src/commands/<noun>/<verb>.ts: the literal ALL_COMMANDS list in packages/cli/src/runtime/command-help.test.ts must include "<noun> <verb>".
For every packages/client/src/domain/<r>.ts: no fixture-and-parse-test pair is required (a unit test of the form Schema.parse(fixture).toEqual(fixture) is a tautology). The schema's contract is tested by the e2e tier when a command consuming it runs against the live API. Report a missing pairing only if no command consumes the schema AND no e2e test parses through it.
For every packages/client/src/resources/<r>.ts: a packages/client/src/resources/<r>.test.ts beside it, covering every exported method with a wire assertion over capture.calls (URL, verb, headers, body) built on captureFetch from packages/client/src/testing/fetch-capture.ts. Report a method the wire test never sends — nothing else pins its path or query vocabulary at the unit tier.
For every pure helper file in packages/client/src/** and in packages/cli/src/{core,output,runtime}/** (excluding index.ts re-exports, the test doubles packages/client/src/testing/fake-client.ts, packages/client/src/testing/fetch-capture.ts and packages/cli/src/core/auth/keyring-mock.ts, and types-only files): a colocated unit test (<name>.test.ts next to <name>.ts, in the same package) exercises it.
For every helper with infinite input space (URL parsing, JSON parsing, projection, byte capping, polling): a property test using should exist.
Subagent 5 — Schema & resource contract, dead exports
Audit packages/client/src/domain/, packages/client/src/resources/, and the export graph across both packages in the current repository. Read-only.
Domain contract — for every packages/client/src/domain/<r>.ts, verify it exports both of:
<Resource> — a z.object({...}).loose() for the full API shape, with type via z.infer. (Zod 4 — .passthrough() is deprecated; .loose() is the project standard.)
<Resource>Compact — <Resource>.pick({...}).strip() with inferred type. The trailing .strip() is mandatory: .pick() on a .loose() parent inherits the loose catchall and silently leaks every API field through the projection at parse time. Report a Compact missing .strip() as path:line: compact-missing-strip: <name>.
Presentation is CLI-owned: the <resource>View (compactPick: <Resource>Compact, tableColumns: ColumnDef[]) belongs in packages/cli/src/output/views/<r>.ts. A ResourceView, ColumnDef, tableColumns, or terminal-presentation format: helper found under packages/client/src/domain/ is a layering violation — report as path:line: presentation-in-client: <name>.
Report missing exports as path: missing-export: <Name>.
Resource contract — for every packages/client/src/resources/<r>.ts, verify it exports <r>Resource(transport), that packages/client/src/client.ts composes the namespace onto the returned object, and that every method follows the eight conventions:
client.<resource>.<method>(...), namespace named after the API resource (mb.card, singular). The resource is what Metabase calls the thing, not what it calls the route: the file and the namespace are snippet even though the endpoint is /api/native-query-snippet, because the domain schema, the CLI noun and the Metabase concept all read snippet.
Path parameters positional, then params, then options: .
Audit packages/client/src/ and packages/cli/src/ in the current repository for correctness and readability smells. Read-only.
Find:
Silent error swallows. Empty catch {} blocks. try/catch whose handler returns a default value ({}, [], "", null) instead of rethrowing or returning a typed error.
Placeholder fallbacks for absent state.?? "", ?? 0, ?? [], ?? {} used to satisfy a type when the real semantic is "missing" or "error" rather than a genuine default. Inspect the surrounding context to judge — a literal default that's user-visible is fine; a coercion that hides a missing API field is not.
Boundary parsing without validation.JSON.parse(...) whose result is consumed without a Zod schema or type guard. Network responses, file contents, env vars treated as already-typed.
Magic literals duplicated across files. Numeric or string constants (timeouts, byte caps, exit codes, profile/account names, file modes like 0o600, retry counts) appearing in 2+ files without a shared named constant. Run a ripgrep pass over likely culprits and report any literal with 2+ occurrences in distinct files, including pairs that straddle the two packages — a constant duplicated across the boundary belongs to whichever package owns the concept, exported for the other.
Boolean trap signatures. Functions with 2+ boolean parameters; call sites passing 2+ positional booleans (foo(true, false)) where the call is not self-documenting.
Type guards that don't validate the narrowing.function isFoo(value): value is Foo whose body checks a property weaker than what distinguishes Foo (e.g., narrowing to NodeJS.ErrnoException on the basis of instanceof Error alone). Quote the guard body.
Unreachable branches. Code paths that cannot fire because an earlier branch returns first, or guards on conditions ruled out at the boundary.
For each, report path:line: <category>: <snippet>. If zero findings, output (clean).
Subagent 7 — Comment & file-size hygiene
Audit packages/client/src/ and packages/cli/src/ in the current repository for comment and size hygiene. Read-only.
Find:
Comments that describe WHAT the code does instead of WHY. Heuristic: if removing the comment would not confuse a future reader who can read the identifier, it's a WHAT-comment.
Comments referencing tasks, PRs, issues, external repo paths, planning documents, dates, or "added/removed/moved by …".
Source files larger than 300 lines.
Single functions larger than 60 lines (count statements between { and matching }; ignore comments and blank lines).
For comments report path:line: <snippet>. For size report path: <LOC> lines and path:fn-name: <LOC> lines. If everything is within bounds, output (clean).
Subagent 8 — Code that fights its tools
Audit the current repository for complexity that exists because the author didn't use the tools they had. Read-only. Not a bug hunt — the code probably works. A leverage hunt: places where bespoke machinery does the job of one idiomatic framework / standard-library / project-helper call.
The class of mistake: the author solved a problem from "what I can build" instead of "what is already available." Eight sub-classes:
Library underuse — reimplementing something the library exports (custom argv parsing, manual subcommand walking, hand-rolled retry, bespoke debouncing, homegrown error hierarchy that the framework already provides).
Library misuse — wrong primitive from the right library; low-level escape hatch when a high-level helper exists; deep imports (/dist/, /internal/, /lib/); detecting library types by string name instead of via exported guards.
Replacing a whole component to alter one seam — wholesale forks of vendor functions, custom dispatch loops, replacement entry points, when a thin intercept at one of the project's own existing seams would do.
Solving at the wrong layer — fixing the symptom at the boundary instead of asking "where else in our own code does this call path already pass through?" If a wrapper / middleware / factory already covers every call site, the fix belongs there.
Standard-library reinvention — manual byte counting, custom URL parsing, hand-rolled deep-equal, ad-hoc base64, homegrown event emitters when node: modules / web platform APIs do it correctly.
Defensive over-engineering for impossible states — try/catch that re-wraps the same error type; validation of trusted internal values; fallback branches that cannot fire; type guards on properties already guaranteed upstream.
Project-helper bypass — direct calls to a primitive when the project owns a wrapper for it. Each direct usage that bypasses the wrapper is drift.
Configuration in code — manual setup that the library accepts as an options bag, plugin, or middleware registration. If the library exposes a hook and the project replaces a function instead, that's almost always the wrong layer.
Phase 1 — surface candidates (each hit is a candidate, not a finding):
Once all eight subagents return, produce ONE report:
# Architectural Audit — <YYYY-MM-DD>
## Summary
- Total findings: <N>
- Critical: <A> High: <B> Medium: <C> Low: <D>
- Clean axes: <comma-separated list of axes that returned (clean)>
## Findings
### Critical
- <path:line> — <rule violated> — <snippet>
...
### High
- ...
### Medium
- ...
### Low
- ...
## Recommendations
1. <concrete next action, ordered by leverage>
2. ...
Severity rubric
Critical — Type-system bypass (casts, any, Record<string,unknown> for typed values, !, @ts-*). Boundary violations (process.exit, raw fetch, JSON.parse, console.*) outside permitted homes. Layering inversion. Silent error swallows in boundary code (catch {}, parse-and-default). Type guards whose body doesn't validate the narrowing. Replacing a whole vendor primitive when an existing project seam covers every call site. Anything that hides drift from the type-checker or the architecture.
High — Missing test pairings. Domain schemas missing required exports. Dead exports. Placeholder fallbacks (?? "" etc.) for absent state. Magic literal duplication across files. Boolean trap signatures. Library underuse / misuse where a documented export does the job in fewer lines (estimated delta ≥ −20 LOC).
nowhere under commands/ — a command calls client.<resource>.<method>(…)
The last two are absolute: neither has a sanctioned home anywhere in the CLI. Report every offending file.
Two rules bind both packages but come from elsewhere, so apply them repo-wide rather than per-package:
console.{log,warn,error,info,debug} — forbidden everywhere, including output/** and tests/. .oxlintrc.json sets no-console: error with no overrides. Report every occurrence.
Platform env vars (XDG_CONFIG_HOME, APPDATA, npm_config_prefix) are read where they are used — core/paths.ts, core/install-method.ts, core/auth/temp-config-home.ts. readEnv owns the MB_/METABASE_ family only; do not report the others.
For each violation, report path:line: <rule>: <snippet>. If zero violations, output (clean).
fast-check
For each missing pairing, report <src-path> → expected <test-path>. If all pairings are present, output (clean).
update(id, params, options?)
Params use Metabase's own field names verbatim — f, model_id, include_inactive. No mapping layer, no renaming.
Transport concerns (signal, timeoutMs, retries) live in the trailing options, never in params.
Wire envelope schemas are module-private to resources/; they never appear in domain/.
Methods return domain values, never wire envelopes. Non-paginated lists return ListResult<T> = { data, total }.
A string path parameter always goes through encodeURIComponent.
Every method carries the endpoint's description as a doc comment.
Report as path:line: <convention-N>: <method-signature-or-path>. Convention 7 is the one with a silent failure mode, so check every ${…} in a path template and confirm the interpolated value is either a number or escaped. Report a resource file no line in client.ts names as path: unreachable-namespace: <r>.
List windowing — for every packages/cli/src/commands/<noun>/list.ts, verify the envelope comes from one of the three helpers in packages/cli/src/output/window.ts, picked by who applied the window: windowList(data, ctx.range, total) when the method returned the whole result set as a ListResult<T> and the slice is client-side (the common case); windowServerPage(data, total, ctx.range) when the endpoint applied limit/offset itself and reports a count; collectForOutput(source, view, ctx) for a genuinely paged endpoint, where source is a PageSource<T> forwarding the helper's PageRequest (max, pageSize) into the resource's <thing>Pages method. An envelope object literal assembled by hand is a drift finding — it produces a plausible shape whose has_more/next_offset no test catches. Report as path:line: hand-rolled-envelope: <command>. Report a source that drops max or pageSize as path:line: unbounded-walk: <command> — the byte budget cannot stop the pull. Also report a list verb that omits ...listFlags, or that never reads ctx.range, as path:line: unwired-range: <command> — its --limit/--offset silently do nothing.
Dead exports — find any export in either package's src/ that is not imported anywhere in packages/*/src/ or tests/. Exclude the entry points (packages/cli/src/cli.ts, packages/cli/src/main.ts) and the client's public barrel packages/client/src/index.ts, whose whole job is re-export. A client export consumed only by the CLI (or only by a test) is live, not dead — resolve @metabase/client/<path> specifiers to packages/client/src/<path>.ts before judging. Report as path:line: dead export: <name>.
Then enumerate the project's own boundary helpers and grep for direct calls to what each helper wraps. Most of them are client surface the CLI is expected to reach for by specifier — parseJson (@metabase/client/json), pollUntil (@metabase/client/poll), paginatePages (@metabase/client/paginate), normalizeUrl (@metabase/client/url), errorMessage / isFileNotFoundError (@metabase/client/errors) — so a CLI file hand-rolling one of those is the highest-yield finding in this axis. The endpoint seam is the client's own: a command that assembles a request from requestParsed (@metabase/client/http/transport) or paginatePages instead of calling client.<resource>.<method>(…) is the same finding one layer up, and the simpler form is a method in packages/client/src/resources/<r>.ts. The CLI's own seams are defineMetabaseCommand (commands/runtime.ts), defineCommandGroup (commands/group.ts), readEnv (core/env.ts), readInput / readBody (runtime/), windowList / windowServerPage / collectForOutput (output/window.ts), and renderItem / renderList (output/render.ts).
Phase 2 — investigate (a candidate must survive all five questions to be promoted):
What is this code's job, in one sentence? If you can't say it, drop the candidate.
What library / framework / runtime is in play?
Does that library/runtime expose a documented way to do this directly? Cite the specific export by name (X from package Y). "The library probably has it" is not a finding.
Does the project's architecture provide a seam here that already wraps every call site? Read importers and their importers — if every call path goes through one chokepoint, the bespoke leaf code is misplaced.
What is the load-bearing reason the bespoke code exists? If you can articulate one (library genuinely cannot do this; seam doesn't actually cover this case; older version had a constraint), the complexity is justified — drop the candidate.
Phase 3 — verify before reporting:
Read the proposed seam — confirm it actually wraps every call site.
Read the cited export — confirm its real signature. If you cite an unexported helper, the finding is dead.
Spot-check whether anything else in the codebase already uses the better pattern. If yes, this is an outlier waiting to be aligned.
Output — one block per finding:
path:line — <category> — <one-line summary>
What's there: <2–3 sentences, quote one short snippet>
Solving: <one sentence>
Simpler: <name the specific export, hook, helper, or seam, and where the small change goes>
Real, not taste: <one sentence — usually "the library exports X" or "every call path already goes through Y">
Delta: <LOC removed minus LOC added — e.g., "−72 / +6">
Verified: <what you read — file+line of the export, importers of the seam>
If zero findings survive Phase 2, output (clean — surveyed N candidates, none survived investigation). State N. Padding with hedged candidates is forbidden.
Forbidden in this audit:
Recommendations that aren't 1–10 lines at an existing seam ("rewrite using framework X" is out of scope).
Findings without a cited export by exact identifier.
Speculation about future framework versions.
Conflating "I would have written this differently" with "the toolset offers a one-liner for this."
Recommending a project helper that doesn't exist yet — that's a separate proposal.
