CLI

anneal audit <file> runs the whole pipeline from the terminal: static analysis, the multi-LLM cascade, the Arsia gas profile, the corpus match, and (optionally) an on-chain attestation.

Run an audit

# No install — run the published CLI straight from npm
npx @tryanneal/cli audit ./Vault.sol --no-llm

# Static only — no API keys, fully deterministic
pnpm --filter @tryanneal/cli start audit ./Vault.sol --no-llm

# Full cascade — ChainGPT pre-screen → Groq + GPT-OSS critics (default; --quick opts out)
CHAINGPT_API_KEY=… GROQ_API_KEY=… \
  pnpm --filter @tryanneal/cli start audit ./Vault.sol

# CI gate — exit non-zero if the verdict scores below 80
npx @tryanneal/cli audit ./Vault.sol --no-llm --threshold 80

# Audit + post the verdict on-chain (mainnet, agent #131)
DEPLOYER_PRIVATE_KEY=0x… \
  pnpm --filter @tryanneal/cli start audit ./Vault.sol --network mantle --attest

Each run opens with an animated TRYANNEAL banner, then prints an is_this_safe() → SAFE / UNSAFE verdict line with the 0–100 score, the list of deduplicated findings (each tagged with the engines that flagged it), and the Arsia gas profile.

A single-contract audit runs the full critic cascade by default — ChainGPT pre-screen, then two architecturally-distinct Stage-2 critics fan out in parallel: Groq Llama-3.3-70B and OpenAI GPT-OSS-120B (both served on Groq), cross-validating each other. Gemini 2.5 Pro is an optional third critic, off by default (its key is rate-limited). Pass --quick for a pre-screen-only pass. The cascade is resilient: a ChainGPT pre-screen failure is non-fatal and the critics still run, and if nothing could analyze the contract the verdict is flagged analysisIncomplete — it is never reported as safe.

Deterministic, reproducible audits.The same contract always returns the same verdict — TryAnneal's answer to “AI audits are non-deterministic.” Every model decodes at temperature 0 (greedy, seeded); a corroboration rule requires every reported finding to have ≥2 independent sources (≥2 models, or a model plus Slither) when the full panel runs, so no single-model hunch drives the verdict; scoring is confidence-weighted; and the Telegram bot and hosted MCP memoize by code hash (keccak/sha3 of the source), so identical source returns the identical audit.

Set HUNYUAN_API_KEY to translate the finished verdict and findings into the reader's language (zh, es, ja, ko, fr, and more) — the audit runs in English, then Tencent Hunyuan renders the multilingual report.

Flags

Exit code is non-zero when a high/critical finding is present, or — with --threshold <score> — when the verdict scores below N (use --threshold 0 for severity-only gating). That single exit code is the whole CI story.

Use it in CI — a GitHub Action PR gate

.github/workflows/anneal-audit.yml runs the deterministic audit (Slither + 16 TryAnneal detectors + the 98-pattern corpus — no LLM keys, no chain calls) on every PR that changes a .sol file. It does three things:

1. Audits each changed contract with --threshold $ANNEAL_THRESHOLD. 2. Posts a ✅ TryAnneal — PASSED / ❌ TryAnneal — BLOCKED comment on the PR with the full per-contract verdict. 3. Emits a red/green check-run that fails when a contract has a high/critical finding or scores below the threshold — so branch protection can block the merge. The threshold defaults to 80 and is set per-repo via the ANNEAL_THRESHOLD repository variable.

# .github/workflows/anneal-audit.yml
name: TryAnneal Security Audit
on:
  pull_request:
    paths: ["**/*.sol"]

permissions:
  contents: read
  pull-requests: write

jobs:
  audit:
    runs-on: ubuntu-latest
    env:
      # Merge gate: block the PR if a contract scores below this (default 80).
      ANNEAL_THRESHOLD: ${{ vars.ANNEAL_THRESHOLD || '80' }}
    steps:
      - uses: actions/checkout@v4
        with: { fetch-depth: 0 }
      - uses: actions/setup-node@v4
        with: { node-version: "22" }
      - uses: actions/setup-python@v5
        with: { python-version: "3.12" }
      - name: Install Slither + detectors
        run: |
          pip install tryanneal-detectors solc-select
          solc-select install 0.8.24 && solc-select use 0.8.24

      # Deterministic gate — exits non-zero on high/critical or score < threshold.
      - name: anneal audit
        run: |
          for f in $(git diff --name-only "origin/${{ github.base_ref }}...HEAD" -- '**/*.sol'); do
            npx @tryanneal/cli audit "$f" --no-llm --no-encrypt --threshold "$ANNEAL_THRESHOLD"
          done