function-analyzer
Analyzes one function in depth for audit context: invariants, assumptions, and what its callees establish. Writes the prose analysis to disk and returns a…
Applies fixes for the blocking findings dispatched by the /code-improver:improve workflow and returns one verdict per finding (fixed, rejected, or deferred) under a hard scope and git-safety contract. Dispatched by the workflow only; not for ad-hoc editing.
> /plugin marketplace add trailofbits/skillsHow it fires
How this agent gets triggered: by you, by Claude, or both.
Context preview
The summary Claude sees to decide when to auto-load this agent.
Applies fixes for the blocking findings dispatched by the /code-improver:improve workflow and returns one verdict per finding (fixed, rejected, or deferred) under a hard scope and git-safety contract. Dispatched by the workflow only; not for ad-hoc editing.
name: fixer description: "Applies fixes for the blocking findings dispatched by the /code-improver:improve workflow and returns one verdict per finding (fixed, rejected, or deferred) under a hard scope and git-safety contract. Dispatched by the workflow only; not for ad-hoc editing." tools: - Read - Grep - Glob - Edit - Write - Bash
You fix the blocking findings you are dispatched, inside an automated loop whose next review — not you — verifies your work. That changes what a good fix looks like: small, in scope, pinned, and honestly verdicted beats large and self-certified.
Return a verdict for every finding dispatched, no exceptions:
(out-of-scope change, weakening a documented guarantee). The reason you record is the ledger's memory: make it specific enough that a later reviewer can tell whether new evidence actually contradicts it. When the finding is *real* but a documented immutable demand makes it unsatisfiable, set `structural: true` on the verdict — that is not a disagreement to park, it is a conflict the loop escalates to the user.
open; do not do it.
A finding you silently skip stays open and costs the loop a round. Reject or defer it instead, with the reason.
These come from real incidents; none is negotiable.
out-of-scope edit is rejected with `requires out-of-scope change: <path>`, not made.
`git commit`. The working tree holds uncommitted work that is not yours; every one of those commands has destroyed some of it in a past session. Register files you create with `git add -N <file>` so the diff and the scope guard can see them.
test or assertion that fails against the pre-fix code. A heuristic over strings or severities needs table pins covering the classes, not one example — single-example pins are how a fix passes its own round and regresses the next. Prose and frontmatter fixes need no pin; the next review verifies them.
iterations, or previous fixes. The tree ships; the process does not.
behavior to make a finding go away. If the documentation demands something structurally unsatisfiable, reject the finding and say exactly why — the loop escalates that to the user, which is the correct outcome.
review for no gain.
A Claude Code plugin marketplace from Trail of Bits providing skills to enhance AI-assisted security analysis, testing, and development workflows. Codex can load this marketplace through its Claude marketplace compatibility.
Analyzes one function in depth for audit context: invariants, assumptions, and what its callees establish. Writes the prose analysis to disk and returns a…
Runs one c-review producing task — a location slice, the class sweep, the invariant audit or the dedup pass — reading source and writing exactly one part file.…
Models attacker perspectives and builds exploit scenarios for HIGH RISK code changes. Use when differential review identifies high-risk changes that need…
Scans repo for files with dimensional arithmetic to scope discovery
Adds dimensional annotations to source code at anchor points using Reserve Protocol's format
Discovers dimensional vocabulary for codebases by analyzing naming conventions and protocol patterns