Review Rust dependencies and create/publish cargo-crev package review proofs for the user. Use when the user asks you to review one of their Rust dependencies, audit a crate, or produce a crev proof.
Installation
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.
Review Rust dependencies and create/publish cargo-crev package review proofs for the user. Use when the user asks you to review one of their Rust dependencies, audit a crate, or produce a crev proof.
cargo-crev: reviewing Rust dependencies as an agent
This skill tells you how to review a Rust dependency on the user's behalf
and help them produce a signed cargo-crev package review proof that they can publish
to their proof repository.
When to use this skill
Use this skill when the user asks you to:
review one of their Rust dependencies (or a transitive dep),
audit a specific crate version,
produce a cargo crev review/proof for a crate,
help them catch up on crates that still need reviewing.
Do not use this skill for unrelated security audits, non-Rust code, or
trust proofs for other crev ids — those are separate workflows.
Prerequisites (check these first)
The user has a crev id configured:
cargo crev id current
If this fails, stop and ask the user to run cargo crev id new --github-username <them> first. Do not auto-generate an id.
The user's proof repo has been set up and is reachable.
cargo crev repo dir prints the local path.
High-level workflow
Refresh the user's web of trust. Run cargo crev update once
at the start of the session to pull the latest proofs from every
trusted reviewer's proof repo (it's a shortcut for
cargo crev repo update). Without this, both candidate discovery
(which depends on the status column reflecting current trusted
reviews) and the later cross-check against existing reviews can be
working off stale data, causing you to either review a crate
someone else just signed off on or miss a recent advisory. See
"Refreshing the web of trust" below.
Pick the crate + version to review. If the user names a specific
crate, skip to the next step. Otherwise, use the candidate discovery
step below to produce a ranked list and confirm the pick with the user.
Start a review report file. See "The review report file" below.
Every finding from every subsequent step — positive or negative —
gets appended to this file as you go. This is your scratchpad, your
audit trail, and the source you distill into the final comment.
Locate the crate's local source. Use
cargo crev crate dir <name> <version> to print the path to
the crate source tree that cargo downloaded from crates.io. Record
the path in the report.
External verification. Check that what crates.io shipped matches
the public upstream repository at a specific commit, and that that
commit corresponds to the advertised version tag. See "External
verification" below. Record the outcome — pass, partial, or fail —
in the report. Any unexpected discrepancy is a real finding and
must be raised to the user before proceeding.
Read the code. See "Reviewing the code" below. Append findings
to the report as you go.
Cross-check against existing reviews. Query the local proof
database for any existing reviews of this crate (any version), read
them, and reconcile with your own findings. See "Cross-checking
existing reviews" below. Record the comparison in the report.
Assemble the unsigned proof file. Fill in the YAML fields
(rating, thoroughness, understanding, comment, issues,
advisories, llm-agent). The comment is the full text of the
report file by default — see "Assembling the unsigned proof"
below.
Validate the proof file. Round-trip the unsigned proof through
cargo crev review --import-unsigned-from ... --no-store --no-edit --print-unsigned. This step is mandatory — fix and re-validate
until it passes. See "Round-trip validating the draft" below.
Hand off to the user. The agent does not sign or publish.
Give the user two files — the report and the unsigned proof —
along with the exact command they should run to interactively
review, edit and sign it. See "Handing off to the user" below.
The review report file
Maintain a single markdown file per review, at:
target/crev/reviews/<crate>-<version>.md
(Create the directories if they don't exist.) Start it as soon as a
candidate is picked, before any inspection begins. Append to it
throughout the review — treat it as an append-only journal.
Append, don't rewrite. If you change your mind about something,
add a follow-up note; don't quietly edit earlier entries. The user
should be able to see how your understanding evolved.
Every problem, however small, goes in. Even if you decide it's
not worth reporting in the final proof, it belongs in the report so
the user can sanity-check your judgement.
Record what you did not look at. If you skipped the test
directory, or only skimmed build.rs, say so explicitly under
"Open questions / things skipped". This is what backs honest
thoroughness / understanding values later.
Hand the report to the user alongside the unsigned proof file
at the end of the workflow (see "Handing off to the user"). The
report is how the user actually audits your work; the unsigned
proof is the thing they sign after auditing.
Refreshing the web of trust
Before any candidate discovery or cross-checking, pull the latest
proofs from every trusted reviewer:
cargo crev update
This is a shortcut for cargo crev repo update. It walks every proof
repository in the user's WoT and git fetches it, so the local proof
database reflects the current state of the world. It can take a few
seconds to a minute depending on how many ids the user trusts and
network conditions; that's normal.
Run this once per session, before the first candidate-discovery
pass, and again only if the session has been running long enough that
the data is plausibly stale or the user explicitly asks. It is not
necessary to re-run it between successive reviews in the same session.
If the command fails for some repos (network blip, dead remote,
auth issue) but succeeds for others, that is usually fine — note the
failures in case they matter, but proceed. If it fails entirely
(e.g. no trusted ids configured, so there is nothing to fetch), stop
and tell the user: candidate discovery and cross-checking will both
be meaningless until the WoT is set up.
Finding candidate crates to review
When the user hasn't named a specific crate, first make sure the WoT
is fresh (see "Refreshing the web of trust" above), then capture a
stats dump of their dependencies.
cargo crev verify is slow — it walks every dependency, downloads
metadata from crates.io, and computes trust graphs. On a non-trivial
project it can take 30 seconds to several minutes. Therefore:
Run it exactly once per session and save the output to a file.
Never re-run it unless the user explicitly asks or the dependency
tree has changed (e.g. after cargo update).
All candidate discovery, sorting, and filtering works on the saved
file — re-read the file instead of re-running the command.
Only request the columns you actually need for candidate selection:
reviews, downloads, loc, and flags. The default output
already includes status, name, and version.
Do not use --show-all — it enables many extra columns (geiger,
owners, issues, lpidx, latest trusted version, …) that produce
wide, hard-to-parse output and are not needed for picking
candidates. If you need a specific extra column later (e.g.
--show-geiger true for a particular analysis), add it to a
one-off query for that crate, not to the full dump.
--force-print-header makes the tool print the column header even
when stdout is redirected. Always use this when capturing — the
header is the only reliable way to know which column is which.
Exit status 255 is expected and does not mean the command
failed. It signals VerificationFailed — i.e. at least one
dependency is not yet fully verified, which is obviously the case
when you're about to start reviewing. Only treat it as a real error
if the stdout file is empty or the stderr contains an actual error
message.
Status column values you'll see:
local — crate from a local path source (skip; not reviewable via
crates.io).
none — no existing trusted review; prime candidate.
pass — already has sufficient trusted reviews; skip.
flagged / dangerous — has issues reported against it; these
may be worth confirming rather than reviewing fresh.
N/A — there are no trusted ids in the user's WoT at all. If you
see this, stop and tell the user they should set up trust first
(cargo crev trust <id>) before candidate discovery is meaningful.
Once the file exists, work from it for all subsequent candidate
selection in this session. If you need to re-examine the data, read
target/crev-verify.txt — do not re-run cargo crev verify.
Before narrowing down, confirm the criteria with the user — see the
"Picking candidates" section below.
Picking candidates
Picking what to review is inherently a bit arbitrary, but the default
goal is: find the non-passing, non-local, least-reviewed,
least-downloaded package — i.e. the most obscure crate in the user's
dependency tree that nobody else is likely to be reviewing any time
soon. Unless the user says otherwise, apply these filters/rankings in
order:
Filter by status. Keep only status = none. Skip pass
(already covered), local (not from crates.io), and N/A
(configuration problem, handled separately). flagged / dangerous
are a separate workflow (confirming pre-existing issue reports) —
don't pick them unless the user asked.
Sort by lowest total review count. Use the second reviews
column (total across all versions, not the per-version count).
Zero-review crates come first.
Tiebreak by lowest total download count. Popular crates will
eventually get reviewed by someone else; obscure ones won't.
Reviewing the long tail is higher marginal value. Use the second
downloads column (total).
Tiebreak by CB flag. Among otherwise equal candidates, prefer
crates with a custom build script — they run arbitrary code at build
time and are higher-risk, so the review is more valuable.
If the user has additional or different preferences (e.g. "I only care
about direct deps", "prioritize high lpidx", "skip -sys crates"), they
override the defaults above — ask if unsure.
The user's own crates are valid candidates. If a top candidate turns
out to be authored by the user themselves, don't filter it out — a
review where the author says "I wrote this and had an LLM agent audit
it" is still a useful signal for downstream consumers, and once
published it'll flip to status = pass and drop out of the filter
naturally. Agent provenance is recorded structurally via the
llm-agent: field on the proof (see "Disclosing agent provenance"
below), so no special comment wording is needed for own-crate reviews.
Pick one candidate, not a shortlist. Filter the captured stats file
by the rules above, sort by the priority order (status → review count →
download count → CB flag), and propose the single most promising crate
to the user. Don't offer a menu — it just slows things down. After that
crate is reviewed and published, repeat the selection process on the
same captured file (or re-capture if it's stale) to pick the next one.
Column legend
These are the columns present in the recommended cargo crev verify
output (with --show-reviews, --show-downloads, --show-loc,
--show-flags). Run cargo crev verify --help for the full legend
including columns not shown by default.
- status - Trust check result: `pass` for trusted, `none` for lacking reviews,
`flagged` or `dangerous` for crates with problem reports.
`N/A` when crev is not configured yet.
- reviews - Number of reviews for the specific version and for all available
versions (total). Two sub-columns: version count / total count.
- downloads - Download counts from crates.io (version / total).
Two sub-columns: version count / total count.
- loc - Lines of Rust code
- flgs - Flags for specific types of packages
- CB - Custom Build (runs arbitrary code at build time)
- UM - Unmaintained crate
- crate - Crate name
- version - Crate version
External verification
Before reading a single line of the crate's code, verify that what
crates.io shipped actually matches the public upstream repository. This
catches the single highest-leverage class of attack (a publisher whose
account was compromised, who can ship a tarball that doesn't match
their public git history). It also tells you which upstream commit
the reviewed snapshot corresponds to, which is valuable information in
its own right.
Ideal outcome, recorded in the report file:
Package matches public repo <url> at commit <sha>, which is the
commit tagged as <tag> corresponding to version <version>.
Anything short of that is a partial or failed verification, and every
discrepancy must be recorded in the report and raised to the user.
Don't silently proceed past a verification issue — if the user hasn't
told you what to do in that case, stop and ask.
Procedure
Get the local source path.
cargo crev crate dir <name> <version>
Record it in the report. Call this path $SRC.
Look for .cargo_vcs_info.json at $SRC/.cargo_vcs_info.json.
When present (it usually is), it contains the upstream git sha the
tarball was built from:
Treat this as a hint, not a guarantee. The file may be missing
(older crates, unusual publish flows), or the sha may be slightly
off (dirty worktree at publish time). If it's missing, note that in
the report and fall back to searching the upstream repo by version
tag instead.
Find the upstream URL. Read $SRC/Cargo.toml.orig (the pristine
pre-publish manifest) and extract package.repository. If absent
or dead, also check package.homepage, then crates.io metadata.
If none of those point at a reachable repository, search the
web (e.g. "<crate-name>" site:github.com or the author name
plus the crate name) — crates sometimes move or get renamed and
the manifest stays stale. Record every URL you tried and which one
resolved. If you still cannot find a public source after a
reasonable search, record the failure and ask the user how to
proceed. Do not skip external verification silently.
Clone (or update) the upstream cache. Use a stable cache
directory so repeat runs don't re-clone:
--filter=blob:none keeps the clone fast even for huge histories —
blobs are fetched on demand when you actually diff.
Identify the upstream commit to compare against. Prefer, in
order:
The sha from .cargo_vcs_info.json (if it exists upstream).
The commit that the version tag points to — tag naming
conventions to try:
v<version>, <version>, <crate>-v<version>, <crate>-<version>.
If neither matches, report the failure and ask the user.
Record in the report which method succeeded and which commit/tag
was chosen. Ideally both the sha and the version tag agree — if
they disagree (e.g. vcs_info points at a commit the tag does not),
that is itself a finding and must be reported.
Check out the commit.
git -C "$UPSTREAM" checkout "$SHA"
Compare the trees. Compare $SRC against
$UPSTREAM/<path_in_vcs> (or $UPSTREAM if path_in_vcs is empty
or .). Use diff -r --brief for an initial pass, then drill into
any differing file with a full diff to see the content.
Files expected to differ or to exist on only one side — add
these to an ignore list and do not flag them:
Cargo.toml — always rewritten by cargo publish. Compare
$SRC/Cargo.toml.orig against upstream's Cargo.toml instead.
Workspace-inherited fields may differ; those differences are
expected and benign as long as they're clearly inheritance
artefacts.
Cargo.toml.orig — exists only in the crate tarball.
.cargo_vcs_info.json — exists only in the crate tarball.
.cargo-ok — cargo extraction marker.
Cargo.lock — sometimes present, sometimes stripped.
.gitignore, .gitattributes — often excluded via exclude.
target/ — build artifacts, never in the tarball.
Any other difference is real and must be recorded in the report
with the file path and a short description of what differs.
Cross-check the manifest. Diff $SRC/Cargo.toml.orig against
$UPSTREAM/<path_in_vcs>/Cargo.toml at the checked-out sha.
Expected differences for workspace-published crates: fields like
version, authors, edition, license being inlined from the
workspace root; [workspace] stanza absent from the crate copy;
path dependencies resolved to version specs. Flag anything else.
Record the outcome in the report under
"## External verification", e.g.:
Package matches public repo
https://github.com/foo/bar at commit
abcdef0123 tagged v1.2.3. Manifest differences are limited
to expected workspace-inheritance artefacts (listed above). All
other files bit-for-bit identical.
or, on failure:
Package does not cleanly match upstream. The file
src/util.rs differs from the upstream copy at the commit
recorded in .cargo_vcs_info.json (see diff below). Raising
to user before continuing.
Any verification failure blocks the rest of the review until the user
says what to do.
Internal verification (reviewing the code)
Once external verification has passed (or the user has explicitly told
you to proceed past a failure), start the actual code review. This is
"internal verification": convincing yourself that the code inside the
verified tarball does what it claims to and nothing more.
Mindset
Approach the code adversarially. Assume, as a working hypothesis,
that everything you are reading could be wrong or hostile, and it is
your job to either falsify that hypothesis or document what you found.
This is not paranoia; it's the baseline for a review that is worth
signing your name to.
Concrete rules that follow from this:
Do not trust comments, docstrings, or variable names at face
value. They are not executable. A function called
sanitize_input can do anything; a comment saying "// does not
allocate" is not a proof. Verify claims by reading the code they
describe, and note in the report when a comment or name is
misleading (even innocently).
Do not trust the crate's README, description, or top-level
rustdoc as a guarantee of behavior. They are part of the "claims"
baseline, and the whole point of the review is to check them against
reality. A crate that says "pure parsing, no I/O" and then calls
std::fs::read is a finding regardless of whether the author meant
well.
Do not trust that the author had good intentions. Assume every
line might have been written by someone trying to smuggle something
past you. If you can't explain why a piece of code is there, that is
itself a finding — write it down, don't paper over it with a
charitable guess.
Do not trust that existing reviews (if any) caught everything.
They might have been cursory, or written before a suspicious line
was added. Every line is your responsibility at the version you are
looking at.
Do not trust test files to reflect real behavior. Tests only
cover the paths the author thought to test (or chose to show). A
malicious crate can have a beautiful test suite. Tests are useful
context, not evidence of correctness.
Do not trust that "obvious" code is harmless. Supply-chain
attacks hide in the most boring-looking places (a string constant,
an inlined byte array, a one-line helper). Give every file a real
look, not a glance.
When something confuses you, record the confusion, don't resolve
it by assumption. Unresolved questions go under "Open questions"
in the report. The user can then decide whether to dig deeper or
downgrade understanding.
The goal is not to be uncharitable — most crates you review will be
fine. The goal is that if something is wrong, your default posture
catches it rather than glossing over it.
Step 1: pick a thoroughness level
Do this before opening any file, and record the decision in the
report with a one-sentence justification. The level dictates which
strategy you follow below.
The canonical cargo-crev definitions (from editing-package-review.md)
are framed for humans in terms of "time per file":
high — hour or more per file; formal security review.
medium — ~15 minutes per file; standard focused review.
low — ~2 minutes per file; quick pass for obvious issues.
none — seconds per file; just skimming, or metadata-only.
For an LLM agent, "time per file" doesn't translate directly (you read
fast, but your value is in careful reasoning, not raw throughput). Use
these rough mappings instead, but stay honest: if you only read half
the files or didn't fully think through the unsafe blocks, downgrade.
Default thoroughness by crate size. Estimate the Rust LoC from the
captured stats file (the loc column) or from
tokei <path> --type Rust if you have it. If the user has specified a
thoroughness level, their instruction overrides these defaults.
Rust LoC
Default thoroughness
Rationale
< 300
medium or high
Small enough to read carefully end-to-end.
300 – 1 500
medium
Read everything once, inspect hotspots.
1 500 – 10 000
low
Full read is impractical; focus on red flags and critical paths.
> 10 000
low or none
Only sample; say so explicitly.
Random file sampling for large crates. When a crate is too large to
read every file (roughly > 1 500 Rust LoC), select a random sample of
at least 10 .rs files and review those thoroughly. Use a command like:
find <src> -name '*.rs' | shuf | head -n 10
Every sampled file MUST be reviewed thoroughly in full — read the entire
file, trace data flow, check for red flags, and record findings. Sampled
files MUST NOT be skipped, read partially, or given a superficial scan.
The whole point of sampling is to give a smaller set of files the deep
attention that the full crate cannot receive; a partial review of a
sampled file defeats the purpose entirely. Record which files were
randomly selected and note in the report that the selection was
randomised. The non-negotiable checks (build.rs, unsafe blocks,
proc-macro, deps, FFI) still apply to the entire crate regardless of
sampling.
Non-negotiable, regardless of level. The following items get
examined in every review, even at low or none:
Form a "claims" baseline from crate metadata, then verify against
reality. Before reading code, read:
The crate-level rustdoc (//! comments in src/lib.rs).
Write a 1–3 sentence summary in the report of what the crate
claims to do. Everything else in the review is essentially a
check that the code matches this claim and does nothing beyond it.
Any capability the code exercises that isn't implied by the claim
(network access, filesystem writes, spawning processes, reading
env vars, FFI, unsafe, pulling in unrelated deps) is a finding
worth recording even if it turns out to be benign. The agent's job
is to answer: "does this crate do only what it claims to do?"
build.rs (any custom build script). Read it top to bottom. This
is the single highest-leverage attack surface — it runs arbitrary
code on the user's machine at build time. Flag on sight: network
calls, shelling out to curl/wget/sh, writing outside OUT_DIR,
reading ~/.ssh/, ~/.aws/, ~/.gnupg/, env-var exfiltration,
obfuscated strings, embedded binary blobs. If the crate has no
build.rs, record that positively in the report.
Cargo.toml dependency list. Cross-check every dependency
against the crate's stated purpose. A JSON parser that pulls in
reqwest is suspicious. A pure-math library that pulls in libc is
suspicious. Git/path deps should not survive cargo publish; if you
see one, that's a finding.
Proc-macro crates (crates with proc-macro = true in
Cargo.toml). Like build.rs, these run arbitrary code at compile
time. Read every #[proc_macro*] function.
Every unsafe block. Use rg -n 'unsafe' <src> to enumerate.
For each, read the surrounding code and make a concrete argument
about why it is (or isn't) sound.
FFI surface (extern "C", bindgen, linked C libraries).
Record the name of every C library dep and note what it's used for.
Everything else scales with thoroughness level.
Step 2: strategy per thoroughness level
Each strategy is additive — medium does everything low does, plus
more, and so on.
none
Metadata-only pass. Use only for tiny trivial crates (e.g. a single
type alias) or when you're creating a proof purely to flag issues
(issues: / advisories:), not to endorse.
Checklist:
Read Cargo.toml and Cargo.toml.orig.
Read README.md (if present) and the crate-level rustdoc.
Record the "claims baseline" in the report (what the crate
says it does).
Run the non-negotiable checks above (claims vs reality,
build.rs, deps, proc-macro, unsafe count, FFI surface).
Record LoC, file count, and what you did not read in the
report.
Do not claim positive or strong rating at this level.
low
Quick single pass focused on red flags. Target: roughly one pass
through every Rust file, spending real attention only on high-value
areas.
In addition to the none checklist:
Enumerate all .rs files. Open each one at least once.
For each file, scan for red-flag patterns (listed below). Dwell
on any hit; move on if clean.
Read every unsafe block, every build.rs, every
proc_macro* function in full.
Note anything that would make you uncomfortable using this
crate yourself, even if you can't prove it's wrong.
Explicitly list in the report which files you read in full
vs. read partially (with line ranges) vs. not read.
Do not use the word "skimmed" — it is ambiguous. As an LLM you
either read the content or you didn't; there is no in-between.
If you read the full file but only did a red-flag grep without
careful analysis, say "read in full (red-flag scan only)".
Red-flag patterns to search for:
unsafe, transmute, from_raw, raw pointer arithmetic.
std::env::var, std::env::vars (env var exfiltration).
reqwest, ureq, hyper, curl, raw TCP sockets — in crates that
don't advertise themselves as network libraries.
include_bytes!, include_str! pointing at suspicious paths.
Base64 string literals, hex blobs, large const byte arrays.
Obfuscated string construction (char-by-char, byte-level, xor'd).
unwrap, expect, panic!, unreachable! in code paths that
handle untrusted input.
Conditional compilation (#[cfg(...)]) that behaves differently on
different targets in non-obvious ways.
Drop impls with side effects beyond freeing memory.
Custom serde deserializers that allocate without bounds.
medium
A careful end-to-end read. Target: understand the crate's overall
architecture, read every non-test file in full, and spend deeper
attention on sensitive areas.
In addition to the low checklist:
Read every file in src/ in full, not just scan.
Sketch a mental model of the crate's public API and internal
data flow. Write a 3–5 sentence summary in the report.
Reconcile against the claims baseline. For every capability
exercised by the code (network, filesystem, processes, env,
FFI, unsafe), confirm it is justified by something in the
crate's stated purpose. Record each reconciliation explicitly:
"reads $HOME/.config/foo — justified, crate is a config
loader". Any unjustified capability is a finding.
For each unsafe block, write a one-sentence soundness argument
in the report (not just "looks fine").
For any parser/deserializer, trace how untrusted bytes reach
each unwrap/expect/panic!.
Skim tests/ and doctests to confirm behavior matches claims.
Run cargo tree on the crate and note any dependency that
surprises you.
high
Formal security review depth. Only claim this level if the user has
explicitly asked for it and you have actually gone this deep.
In addition to the medium checklist:
Multi-pass reading: at least one architecture pass, one
line-by-line pass, and one adversarial pass ("if I were
attacking a user of this crate, where would I look?").
For every unsafe block, work out the safety invariants
explicitly (validity, alignment, aliasing, lifetime, thread
safety) and argue them in the report.
For every public API function that takes untrusted input,
enumerate the failure modes and check that none of them can
corrupt internal state or leak secrets.
Cross-check cryptographic code against the Sherlock /
RustSec-style checklist: canonical serialization, determinism,
constant-time where needed, no ambient state, no
locale-dependent parsing on consensus-critical paths.
If the crate has FFI, note which sanitizer build
(ASAN/MSAN/TSAN) would have exercised it, and whether the
upstream CI actually runs one.
Check release engineering: is the upstream using trusted
publishing? Is the maintainer list stable? Any recent account
changes?
You should almost never produce a high-thoroughness review as an
agent. An LLM can do the reading, but the level implies a human-
depth audit. Default to medium at most unless the user has been
very explicit.
Step 3: record everything in the report
As you review, append to the report file continuously. By the end of
the code review phase, the report's "Code review findings" section
should contain:
The chosen thoroughness level and the justification.
A per-file map of "read in full" / "read in full (red-flag scan
only)" / "read partially (lines X–Y)" / "not read" with
reason. Never use the word "skimmed" — be precise about what you
actually did.
The results of every non-negotiable check (build.rs, deps, proc-
macro, unsafe enumeration, FFI surface) — even if the result is
"none present, positive".
Every red-flag hit, with file:line and a short note on whether it's
a real concern or a false positive (and why).
A short architectural summary (at medium and above).
An explicit list of anything you did not look at and why.
The comment field in the signed proof is distilled from this
section — typically a paragraph or two summarising the above in
reviewer-facing language. The full report stays on disk for the user
to audit your work.
Interpreting results into proof fields
thoroughness: the level you chose in Step 1. Do not inflate it.
understanding: honest self-assessment of how well you grasped the
code. If any unsafe block was "I think this is probably fine but
I'm not sure", that is at most medium understanding.
rating:
strong — secure and good for all applications. Rare; avoid
unless the crate is trivial and perfect.
positive — secure, ok to use, possibly minor issues. The
default for a clean review.
neutral — secure but with flaws worth noting. Default if you
found anything non-trivial to mention.
negative — severe flaws, not ok for production.
(the proof format also supports dangerous for "unsafe to use /
possibly malicious" — use via issues / advisories rather than
the rating field.)
issues / advisories: if you found a concrete bug or security
problem, raise it here. Ask the user before doing so — this is the
kind of thing they probably want to review personally.
Disclosing agent provenance (llm-agent)
Every review produced via this skill must include an llm-agent:
field in the unsigned-review YAML. This is a structured, machine-
readable disclosure that the review was produced by an LLM agent, so
downstream consumers can weight it accordingly. It is not optional and
it is not a substitute for the comment — both exist.
Emit the field with:
model: your own model identifier, as precise as you can get it
(e.g. "claude-opus-4-6", "gpt-4o"). If you genuinely don't know
your own model id, use the closest public identifier you do know.
model-version: include if you know a specific version/date string
for your model; omit the key otherwise (it's optional, skipped when
absent).
human-guided: always emit as false by default. The user may
later edit the proof to flip it to true after they have personally
reviewed and verified the content — that's their call, not the
agent's. Do not set it to true yourself even during an interactive
session.
Example block to drop into the YAML:
llm-agent:model:claude-opus-4-6human-guided:false
This applies to every review (own crates, third-party crates, trivial
crates, large crates — no exceptions). The comment field is for the
actual review findings; it does not need to repeat the "LLM-assisted"
disclosure because llm-agent: already conveys it.
Cross-checking existing reviews
After you've formed your own view of the crate but before you
assemble the proof, check whether anyone else has already reviewed
this crate (any version, not just the target) and read what they
wrote. This is a sanity check on your own work, not a substitute for
it — do it last so your own findings are fully formed and you're not
anchored to a prior reviewer's conclusions.
This step assumes the local proof database is fresh. If you haven't
already run cargo crev update this session (see "Refreshing the
web of trust"), do it now — querying a stale database can miss a
recent advisory or sign-off.
# All reviews of any version of this crate:
cargo crev repo query review <crate-name>
# Only the target version:
cargo crev repo query review <crate-name> <version>
The output is one or more package review YAML documents separated
by ---. For each one, note:
Reviewer id and (if you recognise it) who they are.
Version reviewed.
thoroughness, understanding, rating.
issues / advisories raised, if any.
Key points from comment.
Then reconcile with your own findings. Write a short reconciliation
section into the report under ## Cross-check against existing reviews covering at least:
How many existing reviews there are and for which versions.
Agreement: which of your findings line up with prior reviews,
and which prior ratings are consistent with yours.
Disagreement: any prior finding you could not reproduce, any
finding of your own that prior reviews missed, and any
thoroughness / rating gap. Be specific — reference the prior
reviewer's id and the line of their comment you're responding to.
Verdict: does reading the existing reviews change your
proposed rating, thoroughness, or understanding? If yes,
update your draft fields and say why. If no, say so explicitly
(so the user can see you didn't just ignore them).
Special cases:
Zero existing reviews. Record that explicitly
("No prior reviews of this crate at any version."). This is
common for obscure long-tail picks — it's also why the skill
prioritises those crates.
Prior review flagged a security issue on an older version and
you found a related pattern on the current version. Flag it
loudly in your findings and ask the user whether to raise it as
an issue in your own proof, referencing the prior advisory id.
Prior review's rating is substantially more negative than
yours. Re-read the prior reviewer's comment carefully — they
may have seen something you missed. If after reading you still
disagree, say why in the reconciliation, and consider dropping
your own rating one step (e.g. positive → neutral) out of
epistemic humility.
Prior review's rating is substantially more positive than
yours. Stick to your own findings — do not anchor upward
because someone else was charitable. Explain the gap in your
comment.
Assembling the unsigned proof
Write the unsigned review body to:
target/crev/reviews/<crate>-<version>.proof.yaml
(Same directory as the report file, next to it.) The body is a YAML
document with these fields:
kind:packagereviewversion:-1date:"<RFC3339 timestamp>"from:id-type:crevid:<user'scurrentcrevid—from`cargocrevidcurrent`>url:<user'scurrentcrev-proofsURL—samesource>package:source:"https://crates.io"name:<cratename>version:<crateversion>digest:<recursiveblake2bdigestofthecratesource,base64urlno-padding>review:thoroughness:none|low|medium|highunderstanding:none|low|medium|highrating:negative|neutral|positive|strongllm-agent:model:<yourmodelid,e.g.claude-opus-4-6>human-guided:falsecomment:|
<full contents of the report file, indented under the YAML block
scalar — see rules below>
Put comment: last. The comment block is typically hundreds of
lines of inlined report text, and anything written after it is
effectively hidden from the user when they open the file in an
editor during signing. Keeping comment: at the very end means every
small, easy-to-verify field (from, package, digest, review,
llm-agent) is visible without scrolling past the review body. This
also matches how cargo-crev itself serialises signed proofs —
serde_content_serialize! in crev-data always moves comment to
the bottom — so your draft stays close to the canonical form on
both sides of the signing step.
Populating from. Prefer using the user's actual current crev id
rather than a placeholder. Run:
cargo crev id current
The output is one line per own id, in the form:
<id> <url> (current)
Take the <id> and <url> from the line marked (current) and
write them into from.id and from.url. This makes the draft
self-consistent, lets the dry-validate round-trip (see below) show
the real id in its output, and still works correctly when the user
eventually signs (the signing step will re-set from from the
unlocked id, which will be the same value, so it's a no-op).
Fallback placeholder. If cargo crev id current fails (no id
configured — but the prerequisites section says to stop and ask in
that case, so this should not happen in practice) use the stock
43-character all-A placeholder: AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
with url: "https://placeholder.invalid". The YAML parser requires
from.id to be a valid base64url-encoded crev id (43 chars,
alphabet A–Z a–z 0–9 - _), so arbitrary strings like
"placeholder" will fail parsing.
What goes in comment
By default, comment is the entire text of the report file. The
user can audit a self-contained proof without opening a second file,
the full audit trail travels with the signed artifact, and there's no
risk of the agent quietly dropping inconvenient findings during a
"summarisation" step.
Only compress or re-edit the report on its way into comment if there
is a concrete reason — e.g. the report exceeds cargo-crev's
MAX_PROOF_BODY_LENGTH (~32 KB) and would be rejected as-is. If you
do compress, state that explicitly as the first line of comment,
for example:
[COMPRESSED: original report truncated to fit proof body size limit;
full text at target/crev/reviews/<crate>-<version>.md]
…and keep the untouched report file on disk so the user can inspect
the original. Do not compress "just to be tidy" — verbosity is fine,
silent editorialising is not.
Getting the package digest
Use the dedicated crate digest subcommand. It prints exactly the
value that goes into the digest: field of the proof:
cargo crev crate digest -u <name> <version>
The -u / --unrelated flag tells cargo-crev the crate does not
need to be a direct dependency of the current workspace. The output
is a single line — the recursive blake2b digest, base64url-encoded
without padding — which you paste straight into your YAML.
Do not hand-compute the digest: the recursive-digest algorithm and
its encoding are cargo-crev-specific and easy to get wrong.
Round-trip validating the draft (mandatory)
Every proof file must pass round-trip validation before handoff.
This is not optional — never hand the user a proof that hasn't been
validated, and never skip this step even if the YAML "looks correct".
Run the proof through cargo crev review in dry-validate mode:
With this exact combination of flags (no store, no edit, print
unsigned only), the command does not unlock the user's crev id and
does not produce any side effects — it just parses, normalises, and
re-prints the proof body. If the command succeeds and the output
looks structurally sane (all fields preserved, comment intact),
the file is ready for handoff. If it fails, fix the YAML and
re-validate — repeat until it passes. Never hand the user a draft
that doesn't round-trip.
Handing off to the user
The agent does not sign the proof, does not store it in the
local proof repo, and does not publish. Signing is the user's
affirmative act: it is their key and their reputation on the line.
At the end of the workflow, present the user with exactly three
things:
The path to the report file
(target/crev/reviews/<crate>-<version>.md), with a one-line
summary of the outcome (e.g. "verification passed at
; rated positive at medium thoroughness; one minor
finding in src/util.rs").
The path to the unsigned proof file
(target/crev/reviews/<crate>-<version>.proof.yaml).
The exact command to run to review, edit, and sign:
replace the placeholder from with the user's current crev id,
open the editor on the draft so the user can review, correct
anything the agent got wrong, and flip human-guided: true if
they verified it themselves,
sign and store the proof in the local proof repo once the
user saves and exits the editor.
Note that neither --no-edit nor --no-store appears here —
the user is editing (that's the point) and they are
storing. --no-edit/--no-store are agent-side flags for
preparing drafts, not user-side flags for signing them.
Batch handoff: the signing script
When reviewing multiple crates in a session, maintain a single
executable script that collects all the signing commands:
target/crev/sign-all.sh
Create it at the start of the first review (or when the user asks to
review multiple crates) and append to it as each review completes.
The script should:
Have a #!/usr/bin/env bash shebang and set -euo pipefail.
Contain a singlecargo crev review invocation with multiple
--import-unsigned-from arguments — one per completed review.
The command accepts multiple files and unlocks the user's identity
only once for the whole batch:
Use cargo run -- crev review ... if the session is using the
local build of cargo-crev.
Remove lines for crates the user has already signed. When the
user tells you they've signed some, or you can verify via
cargo crev repo query review <crate> <version>, remove the
corresponding --import-unsigned-from argument from the command.
Be executable (chmod +x).
This way the user can run the script once, signing each review
interactively as the editor opens in sequence, with the passphrase
prompted only once. Tell the user the script path after each batch
of reviews completes.
After the user has signed the proofs, if they want to publish to their
public proof repo, that's a separate explicit step on their side:
cargo crev publish
The agent never runs publish.
Rating / thoroughness / understanding cheat-sheet
TODO(interactive): record the user's personal calibration for these
values. For now, default to conservative choices and ask.
rating: start with neutral unless the user tells you otherwise.
thoroughness: reflect how much you actually read.
understanding: reflect how much you actually understood.
Things to never do without asking
Never sign the proof yourself. The agent prepares the unsigned
YAML file; the user runs cargo crev review --import-unsigned-from
to edit and sign. Signing is the user's affirmative act.
Never run cargo crev publish. Publishing is a separate,
explicit step the user takes after they've audited and signed.
Never fabricate a review you didn't actually perform. If you skipped
files, say so in the comment.
Never set thoroughness: high or understanding: high unless the user
has explicitly confirmed.
Never create trust proofs (cargo crev id trust) from this skill —
that's a separate, higher-stakes action.
Never omit the llm-agent: field from a review you produce, and
never set human-guided: true yourself. Flipping that flag is the
user's decision after they personally verify the review.
Never hand off a proof file that hasn't passed round-trip validation
(cargo crev review --import-unsigned-from ... --no-store --no-edit --print-unsigned). Always validate, fix, and re-validate until it
passes.
Delegation policy (sub-agents)
Prefer doing reviews yourself. The review workflow described in this
skill requires careful, adversarial reading of code, nuanced judgement
about unsafe blocks and FFI, thorough external verification with diffs,
and honest self-assessment of thoroughness/understanding. These are
difficult to delegate well.
Default: do not delegate reviews to sub-agents. Perform each review
in the main conversation context where you have the full skill
instructions, the review report as a living document, and can maintain
the adversarial mindset throughout.
If the user explicitly asks you to parallelise (e.g. "review several
at once", "keep going in the background"), you may delegate to
sub-agents, but only under these conditions:
Pass the full skill context. The sub-agent prompt must include
the complete review procedure — not a summary, not bullet points,
but the actual workflow steps, the thoroughness-level checklists,
the external verification procedure, the non-negotiable checks, and
the adversarial-mindset rules from this skill file. If the prompt
would be too long, do fewer reviews in parallel rather than cutting
corners on the instructions.
Include the exact proof YAML template with the user's crev id,
proof URL, and package digest already filled in, so the sub-agent
doesn't have to guess or look them up. Also pass the path to the
already-captured target/crev-verify.txt file and tell the
sub-agent to read it, not re-run cargo crev verify — the
command is slow and has already been run.
Specify the thoroughness floor. Tell the sub-agent which
thoroughness level to target (based on LoC) and that it must not
inflate its self-assessment. If it only skimmed files, it must say
low, not medium.
Require the full report file. The sub-agent must produce the
same structured report (target/crev/reviews/<crate>-<version>.md)
with all sections: external verification, file map with read/skimmed
annotations, non-negotiable checks, red-flag scan results,
architecture summary, claims-vs-reality reconciliation, cross-check
against existing reviews, and open questions.
Quality over quantity. It is always better to produce 3 thorough
reviews than 10 shallow ones. A review that inflates its
thoroughness or misses findings is worse than no review — it gives
false confidence to downstream consumers. If you cannot maintain
quality at the requested parallelism, reduce the parallelism.
Verify sub-agent output. After a sub-agent completes, spot-check
its report before adding the proof to the signing script. If the
report is shallow (e.g. no external verification diff, no per-file
annotations, no unsafe analysis), either redo the review yourself
or send the sub-agent back with corrections.
Open questions for the user (interactive session)
These are placeholders for the user's interactive input. Update this file
as answers come in.
What files/directories do you typically skip when reviewing a crate?
How do you decide between issue and advisory?
How do you pick thoroughness and understanding values?
Do you want agent reviews tagged somehow in the comment (e.g. a
prefix like [agent-assisted])?
Any crates/patterns that are automatic red flags for you (build.rs,
unsafe, proc macros, network calls, etc.)?