Trustabl fix agent reliability

Trustabl scans and fixes your agents for production reliability and performance issues across Claude Agent SDK, OpenAI Agents SDK, Google ADK, MCP, LangChain, LangGraph, CrewAI, AutoGen AG2, Pydantic AI, and Vercel AI.

A GitHub Action that surfaces the results where you work:

• Inline PR annotations + the Security tab. Findings are uploaded to GitHub Code Scanning, so they appear on the changed lines in the PR diff and in the repository's Security tab.
• A sticky PR comment with the readiness score, severity breakdown, and a fix-headroom ladder — updated in place on each run.
• Status-check gating. Optionally fail the job on a risk-score or severity threshold so it can be a required check.
• A readiness panel in the run log and the Step Summary.
• Optional dependency CVE scan (vuln-scan: true) — matches your declared dependencies against a pinned OSV snapshot and reports known CVEs as findings, so they appear on every surface (score, gate, annotations, Security tab).

It downloads the official trustabl release binary (sha256-verified against the release checksums.txt), tool-caches it, scans your checkout, and reports.

Quick start

name: Trustabl
on: [push, pull_request]

permissions:
  contents: read
  security-events: write   # SARIF → Security tab + inline PR alerts
  pull-requests: write     # sticky PR comment

jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: trustabl/trustabl-action@v0

With zero config the action scans the checkout, posts findings to the Security tab and (on PRs) as a sticky comment + inline annotations, attaches trustabl.json + trustabl.sarif as an artifact, and fails the job only if trustabl itself flags a medium-or-higher finding.

Permissions. security-events: write enables the Security tab upload; pull-requests: write enables the comment. If you omit either, the action degrades gracefully — it warns and falls back to inline annotations + the Step Summary, and never fails solely because a surface was unavailable. On fork PRs the token is read-only, so the upload and comment are skipped automatically.

Annotated example

name: Trustabl
on:
  push:
    branches: [main]
  pull_request:
  workflow_dispatch:

permissions:
  contents: read
  security-events: write
  pull-requests: write

jobs:
  scan:
    runs-on: ubuntu-latest            # also works on macos-* / windows-*
    steps:
      - uses: actions/checkout@v4     # REQUIRED first — the action scans your checkout
      - uses: trustabl/trustabl-action@v0
        with:                         # every input is optional
          # detectors: openai_sdk           # limit SDKs: claude_sdk,openai_sdk,google_adk,openshell
          # version: latest                 # trustabl release to run; pin e.g. v0.5.0 for reproducible CI
          # vuln-scan: true                 # also scan dependencies for known CVEs (OSV)
          # severity-threshold: high        # fail if any finding >= level (none|low|medium|high|critical)
          # risk-score-threshold: 70        # fail if risk (100 - readiness) >= N (0 disables)
          # comment-on-pr: true             # sticky PR summary comment
          # annotations: true               # inline annotations
          # upload-sarif: true              # Security tab upload
          # upload-artifact: true           # attach trustabl.json + trustabl.sarif

Pinned + gated

- uses: trustabl/trustabl-action@v0.3.1
  with:
    version: v0.5.0
    detectors: claude_sdk,openai_sdk
    severity-threshold: high       # fail on any high or critical finding
    risk-score-threshold: 70       # fail if risk (100 - readiness) >= 70
    artifact-retention-days: "30"

Enrich + auto-enrich

When enrich: true, after the scan the action calls trustabl enrich with your LLM API key to generate AI explanations and code fixes for each finding. With auto-enrich: true, high-confidence fixes are applied directly to source files. With create-fix-pr: true, the patches are committed on a new branch and a pull request is opened for human review.

permissions:
  contents: write        # push fix branch
  pull-requests: write   # open fix PR
  security-events: write
  
jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: trustabl/trustabl-action@v0
        with:
          enrich: true
          llm-key: ${{ secrets.ANTHROPIC_API_KEY }}
          auto-enrich: true
          create-fix-pr: true

Enrich is best-effort — if it fails the scan result and gate decision are unaffected and a warning is emitted instead of failing the job.

PR-only auto-enrich. To generate explanations on every push but only apply fixes and open a fix PR on pull requests:

- uses: trustabl/trustabl-action@v0
  with:
    enrich: true
    llm-key: ${{ secrets.ANTHROPIC_API_KEY }}
    auto-enrich: ${{ github.event_name == 'pull_request' }}
    create-fix-pr: ${{ github.event_name == 'pull_request' }}

Required repo settings when using create-fix-pr: true: Go to Settings → Actions → General → Workflow permissions and enable Read and write permissions + Allow GitHub Actions to create and approve pull requests.

Inputs

Name	Default	Description
target	.	Path or GitHub URL to scan.
version	latest	trustabl release tag (e.g. v0.5.0) or latest.
detectors	(all)	Comma-separated SDK subset: claude_sdk, openai_sdk, google_adk, openshell, mcp, langchain, crewai, pydantic_ai, vercel_ai, autogen.
strict	false	Pass --strict (fail on any finding).
vuln-scan	false	Match dependencies against a pinned OSV snapshot; report known CVEs as findings.
rules-ref	(default)	Pin a trustabl-rules git ref.
rules-repo	(default)	Override the trustabl-rules source repo.
upload-sarif	true	Upload SARIF to Code Scanning. Needs security-events: write.
sarif-file	trustabl.sarif	SARIF output path.
json-file	trustabl.json	JSON ScanResult output path.
upload-artifact	true	Attach JSON + SARIF as a workflow artifact.
artifact-name	trustabl-scan-results	Artifact name.
artifact-retention-days	(repo default)	Days to keep the artifact (1-90).
comment-on-pr	true	Sticky PR summary comment. Needs pull-requests: write.
annotations	true	Emit inline annotations for findings.
max-annotations	10	Max inline annotations (worst-severity first).
risk-score-threshold	0	Fail when risk >= N (0 disables).
severity-threshold	none	Fail when any finding >= severity (none/low/medium/high/critical).
branch	(auto)	Report branch label; auto-detected from the checkout.
github-token	${{ github.token }}	Token for release lookup, SARIF upload, and PR comments.
enrich	false	Run AI enrichment on findings (explanations + fixes). Requires llm-key.
llm-provider	anthropic	LLM provider for enrichment (e.g. anthropic).
llm-key	(none)	API key for the LLM provider (BYOK). Required when enrich is true.
auto-enrich	false	Apply AI-generated fixes to source files. Requires enrich: true.
create-fix-pr	false	Open a PR with applied fixes. Requires auto-enrich: true. Needs contents: write + pull-requests: write.
enrich-model	(binary default)	Claude model for enrichment (e.g. claude-sonnet-4-6). Defaults to claude-haiku-4-5.
enrich-rules	(all)	Comma-separated rule IDs to enrich (e.g. ADK-201,ADK-105). Empty = all findings.
fix-pr-base	(current branch)	Base branch for the fix PR.

Outputs

Name	Description
exit-code	trustabl native exit code (0 / 1 / 2).
readiness-score	Integer percent [0, 100], higher = better.
risk-score	100 - readiness-score.
max-severity	Highest severity among findings, or none.
findings-count	Total finding count.
sarif-uploaded	Whether Code Scanning accepted the SARIF (true/false).
sarif-file	Path to the emitted SARIF file.
json-file	Path to the emitted JSON file.
artifact-name	Artifact name used for the upload.
enrich-json-file	Path to enriched.json (when enrich is true).
fix-pr-url	URL of the opened fix PR (when create-fix-pr is true).

How it works

• Verified install. The release asset for the runner's OS/arch is downloaded and its sha256 checked against the release checksums.txt before it runs, then tool-cached so reruns skip the download.
• Single scan. When the installed engine supports --json-out/--sarif-out, one analysis pass produces both artifacts. Older engines fall back to two scans automatically (and the headroom ladder is hidden, since it needs the engine's projected_scores). Use version: latest to get the fast path.
• Dependency CVE scan (opt-in). With vuln-scan: true, declared dependencies are matched against a pinned OSV snapshot; each known CVE becomes a finding (so it counts toward the score, gate, annotations, and Security tab), plus a dependencies-scanned / known-vulnerabilities line in every report. The OSV database is fetched once on first use, then cached.
• Honest gating. A failed or empty scan errors the job rather than reporting a clean score. The gate decision is exit-code/threshold-based, surfaced in the Step Summary and the PR comment.

Downloading the scan result

After a run, open the run page and find the trustabl-scan-results artifact:

• trustabl.json — full machine-readable ScanResult.
• trustabl.sarif — SARIF 2.1.0.

- uses: actions/download-artifact@v4
  with:
    name: trustabl-scan-results

Versioning

• Pin a release: uses: trustabl/trustabl-action@v0.3.1.
• Or track the line: uses: trustabl/trustabl-action@v0 (the moving major tag).

Notes

• Runs on ubuntu-*, macos-*, and windows-* runners (x64/arm64; Windows is amd64-only, matching the trustabl release matrix).
• actions/checkout must run first — the action scans your checkout.
• security-events: write is required for the Security tab upload; pull-requests: write for the sticky comment. Both degrade gracefully if absent.

Development

This is a node24 TypeScript action bundled to dist/ with ncc.

npm ci
npm run typecheck   # tsc --noEmit
npm test            # jest unit tests
npm run build       # bundle to dist/index.js (commit the result)
npm run all         # all of the above

dist/ is committed because a node24 action runs dist/index.js directly from the consumer's checkout of the release tag. The Build check workflow fails a PR whose dist/ is stale, so always npm run build and commit after changing src/.