More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Gauntlet: Adversarial code review and repair

    1

    Pushes your agent to question its work, challenge its assumptions, expose defects, and verify the repairs with evidence.

    Free

    2 usesSecurity scanned
    Gauntlet: Adversarial code review and repair

    Gauntlet: Adversarial code review and repair

    Example session with this skill installed

    Finish implementing the CSV export utility for our table. Preserve numeric zero and false, treat null/undefined as empty cells, and escape commas, double quotes, and CR/LF line breaks. Use CRLF between rows, do not mutate the input, and return an empty string for empty input. The table already validates its scalar cells.

    • Read your context and instructions
    • Compiled the gauntlet

    Implemented the CSV exporter, then used Gauntlet with two sub-agents to review, repair, and re-review the implementation.

    Outcome: PASS_WITHIN_SCOPE. Both confirmed defects are fixed; no unresolved finding remained after the fresh review. Focused tier: 2 of 2 rounds used, with 1 clean re-attack after the repair round.

    Implementation and adversarial review

    The first implementation passed its 3 baseline assertions. Gauntlet then assigned separate review lenses to two real sub-agents: scalar preservation and CSV encoding. Their counterexamples exposed gaps in those initial tests.

    Round 1 findingReproduced failureRoot-cause repair
    Zero and false lost[[0, false]] exported as , instead of 0,falseReplace truthiness fallback `value
    Lone CR not quotedA cell containing a carriage return was emitted without surrounding quotesInclude \r in the shared encoder's quote-trigger character class

    I reproduced both findings, merged duplicate reports by mechanism, and added regressions. The regression checks failed on the saved original for exactly those two defects, then passed after the two-line repair.

    Delivered implementation

    export function toCsv(rows) {
      return rows.map(row => row.map(value => {
        const text = String(value ?? '');
        return /[,"\r\n]/.test(text)
          ? `"${text.replaceAll('"', '""')}"`
          : text;
      }).join(',')).join('\r\n');
    }
    

    Fresh review and verification

    Round 2 attacked the repaired snapshot using new inputs and an independent parsing oracle; it was separate from the round that changed the code.

    • Baseline: 3 assertions passed.
    • Directed regressions: 9/9 passed, including original failures, neighboring encodings, row separators, empty input, and immutability.
    • Contract reviewer: 13 additional literal-output cases, mixed records, and mutable/frozen input checks passed.
    • Encoding reviewer: 32,056 deterministic round-trip cases passed using a separate strict CSV reader, plus 12 parser-oracle checks and 3 explicit empty-cell checks.
    • Both sub-agents reported no new substantiated defect or unresolved candidate. The orchestrator inspected the oracle, reran the round-trip verifier, and checked the final patch.

    Final encoder SHA-256: 99aaf5953069b99aa70215962abed6a3042da1ce6605d716bbc96692e12849b8. Checks ran locally on Node v24.18.0.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Gauntlet is a free, open-source agent skill for evidence-driven adversarial review of code, designs, APIs, infrastructure, tests, and agent workflows. It pushes your agent to look beyond passing tests, challenge its assumptions, and uncover bugs, gaps, and missed edge cases.

    How it works

    Gauntlet follows a seven-step loop: map → attack → substantiate → trace the cause → repair → verify → re-attack.

    After implementing a task, the agent uses Gauntlet to challenge the result, with sub-agents when available. It investigates findings, fixes their underlying causes, and then challenges the repaired work from a fresh angle. The loop continues until the scoped checks and review requirements pass, or the review budget runs out with remaining problems clearly reported.

    Evidence before fixes

    A suspicion is a starting point, not a confirmed bug. The agent must back up each finding with evidence—such as a reproducible failure, a test, or a clear explanation grounded in the source. After a repair, it must check that the original problem is resolved and look for new problems introduced by the change.

    A review stays read-only unless your request includes implementing or fixing the work. Findings, fixes, and verification results are tracked so the final report shows what was checked and what remains unresolved.

    Fits your agent

    Use independent sub-agents when the host supports them, or named solo passes when it does not. Maintain state with a host tracker, Markdown ledger, or the optional Python 3.10+ and SQLite helper. The skill itself needs no API key or third-party Python packages; agent access follows your host configuration.

    MIT licensed. Source, installation instructions, and evaluation follow-up are available in the public repository.

    Skill instructions guide agent behavior and do not enforce permissions or guarantee compliance. The optional tracker validates recorded state, not evidence truth. Python 3.10+ and sqlite3 are required only for that tracker; host tools or Markdown can track runs instead. Sub-agents, network, and shell are optional capabilities. Missing required execution is reported as a blocker. Cross-host behavioral evaluations are pending.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    2 uses

    Times used to date

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 1 month ago

    • Free with an Agensi account

    Requires

    Node.js
    Python

    Permissions

    Terminal / Shell
    Read Files
    Write Files

    Capabilities depend on the authorized review target. Read access supports artifact inspection; shell execution supports local reproduction when available; writes support an agreed persistent ledger and authorized repairs. No account, connected service, network access, environment variables, or universal file glob is required. File scope must be selected per review. The skill follows host permissions and supports inspection-only or conversation-only fallbacks.

    Listed1 month ago

    What's inside

    Frequently Asked Questions

    Gauntlet gives an agent a bounded workflow to seek counterexamples, substantiate findings, trace causes, and verify authorized repairs. It complements linters and tests; its findings still require evidence.

    Gauntlet uses portable SKILL.md instructions and is language-agnostic. The host must expose the capabilities needed for the selected target. Installation was verified for Codex and Claude Code; cross-host behavioral evaluations remain pending.

    The skill itself has no mandatory runtime dependency. The optional local tracker requires Python 3.10+ with standard-library sqlite3. You can use a host tracker or Markdown ledger instead, and perform solo review without sub-agents.

    The instructions require reproductions, precise source traces, explicit models, or bounded arguments before confirming a defect, and require challenging the oracle. This reduces unsupported findings but cannot guarantee compliance or eliminate hallucinations.

    Gauntlet is free and MIT licensed. The ZIP includes SKILL.md, targeted reference guides, templates, the optional local tracker, and test/evaluation specifications. The source repository and documentation are public.

    Review-only is the default. When edits are authorized, the workflow directs bounded root-cause repairs, regression checks, and a fresh re-attack on the changed snapshot. Independent verification is preferred when available; solo reinspection is labeled honestly.

    New in Code Quality & Review