| name | exemplar-audit |
| description | Compare implementation against canonical open-source references. Clone exemplars into /tmp, diff edge cases, error paths, and protocol invariants. Use when building a new subsystem, after a refactor, or when reviewing code that implements a known standard (POSIX, FUSE, WAL, S3, NFS, etc.). |
| allowed-tools | Read, Grep, Glob, Shell, WebSearch, StrReplace |
Exemplar Audit
Find what the best implementations do that we don't, then close the
gaps. This is the "measure twice" pass: compare our code against
battle-tested references before declaring it done.
When to use
- After implementing a new subsystem (filesystem, protocol, cache)
- After a refactor that changed invariant-bearing code
- When reviewing code that implements a known standard
- When the user says
/exemplar-audit or /refine --pass 6
- When you suspect edge cases are missing but don't know which ones
Procedure
1. Identify exemplars
Check these sources in order:
- This monorepo first. The best reference for "how we do X" is
often another subsystem that already does X well. GitHub Actions
workflows, Terraform modules, Rust crates, Go controllers -- scan
for the most mature sibling before looking externally.
- Module docs and READMEs in the subsystem being audited.
Look for "References", "Inspired by", "See also", or citations
to external projects.
storage.md, architecture.md, or equivalent design docs.
These often name reference implementations.
- The internet. Search for "canonical open-source implementation
of {thing}". Prefer projects with >1000 stars, active maintenance,
and a test suite.
Examples of good exemplars by domain:
| Domain | Internal exemplar | External exemplar |
|---|
| GitHub Actions | host-agent-rust.yml, host-agent-kvm.yml | -- |
| Terraform | apps/cloud/apps/dcs/tf/main.tf | checkov rules |
| Rust CI crate | host-agent (same workspace) | -- |
| FUSE filesystem | -- | fuser SimpleFS, FUSE memfs |
| SQLite metadata | -- | JuiceFS pkg/meta/sql.go |
| WAL replication | -- | Litestream, libSQL bottomless |
| S3 chunk store | -- | JuiceFS pkg/object/s3.go |
| VFS layer | -- | Linux fs/fuse/dir.c |
| Content-addressed storage | -- | git sha1-file.c, Perkeep |
2. Clone or read
Clone into /tmp to avoid polluting the workspace:
git clone --depth 1 https://github.com/cberner/fuser /tmp/fuser
git clone --depth 1 https://github.com/juicedata/juicefs /tmp/juicefs
For smaller references, read online (GitHub raw, crates.io source).
3. Compare
For each exemplar, check:
- Edge cases they handle that we don't. Look at their test suite
and error paths, not just the happy path. grep for
ENOTEMPTY,
ENOSPC, EINVAL, rename, unlink, etc.
- Error paths they enforce that we skip. Do they validate inputs
we trust? Do they check return values we ignore?
- Protocol invariants they check that we miss. POSIX rename
atomicity, nlink consistency, fsync durability, WAL checkpoint
ordering.
- Patterns they use that are better than ours. Simpler error
mapping, cleaner state machines, fewer allocations.
4. Report and fix
For each gap found:
- State the invariant or edge case in one sentence.
- Cite the exemplar file and line.
- Check if we have a test for it. If not, write one (fail first).
- Fix the code if needed.
Do not cargo-cult patterns we don't need. If the exemplar handles
NFS-specific locking and we don't serve NFS, skip it. Every adopted
pattern must have a real consumer in our system.
What NOT to do
- Don't clone 50 repos and compare superficially. Pick 2-3 good ones.
- Don't adopt patterns because they're clever. Adopt because they
prevent bugs we'd otherwise ship.
- Don't report "they do X and we don't" without checking if X is
relevant to our use case.
- Don't spend more than 30 minutes per exemplar. If the comparison
isn't yielding findings, move on.
Output
Return a table:
| Gap | Exemplar | Our status | Fix |
|---|
| rename over non-empty dir | fuser SimpleFS L412 | missing | added test + check |
| WAL checkpoint past replicated | Litestream wal.go L890 | missing | added watermark guard |
| (none found) | JuiceFS chunk GC | clean | -- |