Skip to content

README.md

Latest commit

 

History

History
201 lines (162 loc) · 6.58 KB

README.md

File metadata and controls

201 lines (162 loc) · 6.58 KB
Raw

Issue Labeler

A reusable GitHub Action that classifies issues using an LLM (via the Strands Agents SDK) and applies labels from a hardcoded allowlist.

Security Model

The LLM has no tools, no shell access, and no GitHub API access. It returns a structured object whose label field is an enum built from the label names in your config file, so out-of-allowlist values are rejected by the schema rather than filtered after the fact. The worst a prompt injection can achieve is mislabeling — never arbitrary label creation or other side effects.

Layer Protection
Input Sanitized (control chars stripped) and truncated before reaching the LLM
Output Structured output constrained to an allowlist enum, capped at max_labels
IAM Scoped to bedrock:InvokeModel / bedrock:InvokeModelWithResponseStream
Permissions Workflow only needs issues: write

Usage

Single labeler (file config)

# .github/workflows/issue-labeler.yml
on:
  issues:
    types: [opened, edited]

permissions:
  issues: write
  id-token: write
  contents: read

jobs:
  label:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with:
          sparse-checkout: .github/labelers
          sparse-checkout-cone-mode: false
      - uses: strands-agents/devtools/issue-labeler@main
        with:
          aws_role_arn: ${{ secrets.AWS_ROLE_ARN }}
          config_path: '.github/labelers/area.yml'

Single labeler (inline config)

- uses: strands-agents/devtools/issue-labeler@main
  with:
    aws_role_arn: ${{ secrets.AWS_ROLE_ARN }}
    max_labels: '1'
    config: |
      labels:
        bug: "Something is broken"
        enhancement: "New feature or improvement"
        question: "User needs help"

Multiple independent labelers

Run separate jobs for each concern. They execute in parallel and don't conflict:

jobs:
  label-area:
    steps:
      - uses: strands-agents/devtools/issue-labeler@main
        with:
          aws_role_arn: ${{ secrets.AWS_ROLE_ARN }}
          config_path: '.github/labelers/area.yml'

  label-type:
    steps:
      - uses: strands-agents/devtools/issue-labeler@main
        with:
          aws_role_arn: ${{ secrets.AWS_ROLE_ARN }}
          config_path: '.github/labelers/type.yml'
          max_labels: '1'

  label-language:
    steps:
      - uses: strands-agents/devtools/issue-labeler@main
        with:
          aws_role_arn: ${{ secrets.AWS_ROLE_ARN }}
          config_path: '.github/labelers/language.yml'
          max_labels: '1'

Config File Format

# Optional: extra context appended to the system prompt
instructions: |
  CI dependency bumps should be labeled "chore".

# Required: label allowlist. Keys are exact label names.
# Values can be a string (description) or an object with a "description" key.
labels:
  area-mcp:
    description: "Model Context Protocol, MCP servers/clients/transport"
  area-provider:
    description: "Model providers (Bedrock, OpenAI, Anthropic, Ollama)"
  # Shorthand form:
  area-tool: "Tool behavior, execution, @tool decorator"

Setting native issue type and issue fields (issues only)

In addition to labels, the action can set the org's native issue type and a single-select issue field, whenever the triggering event carries an issue payload (these features do not apply to pull requests). This is additive: labels are still applied exactly as before.

Add a type: key to a label to map it to a native issue type:

labels:
  bug:
    description: "Something is broken"
    type: Bug          # native issue type name (resolved at runtime)
  enhancement:
    description: "New feature or improvement"
    type: Feature

Add a top-level field: block plus per-label option: keys to set a single-select issue field:

field:
  name: Language       # issue field name (resolved at runtime)
labels:
  python:
    description: "Python SDK"
    option: Python     # single-select option name
  typescript:
    description: "TypeScript SDK"
    option: TypeScript

Names (Bug, Language, Python) are resolved to GitHub node IDs at runtime via a repo-level GraphQL query and matched case-insensitively. A malformed field: block fails the run (like a malformed labels: block); after that, unmatched names or API errors emit a warning and are skipped, and never fail the workflow. Existing values are always overwritten. The action's existing issues: write permission is expected to cover these mutations; verify against a test issue before rolling out broadly.

Inputs

Input Required Default Description
aws_role_arn Yes - AWS IAM role for Bedrock
config No* - Inline YAML config
config_path No* - Path to config file
aws_region No us-west-2 Bedrock region
model_id No global.anthropic.claude-sonnet-4-6 Bedrock inference profile for classification
max_labels No 3 Max labels per issue
max_body_length No 1000 Max chars of body sent to LLM

*One of config or config_path is required.

Outputs

Output Description
labels Comma-separated labels applied (empty string if none)

Prerequisites

  1. An AWS IAM role that the GitHub OIDC provider can assume
  2. The role needs bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream permissions
  3. Bedrock access enabled for the configured model in aws_region
  4. The labels referenced in your config must already exist on the repo

Backfilling existing issues

backfill.py applies the same native type/field logic to existing issues. For each issue it reuses an existing type/language label when present, and only calls the LLM for issues missing one; freshly classified labels are applied to the issue so re-runs reuse them. The run aborts up front if the configured type/field/option names don't resolve on the repo, and exits non-zero listing any issues whose updates failed. Run locally with gh authenticated and AWS credentials available:

pip install --upgrade strands-agents pyyaml "boto3>=1.35.0"

# Dry run first (prints intended changes, writes nothing):
python3 backfill.py --repo strands-agents/harness-sdk \
  --type-config path/to/type.yml \
  --language-config path/to/language.yml \
  --dry-run

# Apply (all issues, open and closed, by default):
python3 backfill.py --repo strands-agents/harness-sdk \
  --type-config path/to/type.yml \
  --language-config path/to/language.yml

For Python-only repos (e.g. evals) omit --language-config.