Skip to content

issue-bot.md

Latest commit

 

History

History
75 lines (61 loc) · 4.63 KB

issue-bot.md

File metadata and controls

75 lines (61 loc) · 4.63 KB
Raw

Hosted issue bot

English | Türkçe

The hosted bot lets a visitor scan one public GitHub repository from a structured issue form. It is a convenience layer around the same deterministic scanner used by the CLI.

Request flow

  1. Select Herkese açık depoyu tara / Scan a public repository on the new-issue page.
  2. Choose Türkçe or English as the report language.
  3. Enter exactly one https://github.com/OWNER/REPOSITORY URL on its own line.
  4. Submit the issue. The form adds the scan-request label.
  5. The bot resolves the target's default branch to an immutable commit SHA.
  6. A deterministic report and, when available, an AI-generated explanation are posted in the selected language as one bot comment.

Reopening the issue reruns the scan and updates the existing bot comment when it is among the first 100 comments.

Deterministic scans run automatically for every valid form submission. To protect the anonymous provider quota from public issue spam, AI explanations run automatically only when the requester's association with this WorkflowPromptGuard repository is OWNER, MEMBER, or COLLABORATOR. A maintainer can enable AI for another request by applying the ai-approved label.

Authentication and cost

No third-party key or long-lived repository secret is required. GitHub creates short-lived GITHUB_TOKEN credentials for GitHub operations: the scan job performs read-only GitHub API requests, and the comment job uses issues: write. The model request is sent anonymously to https://api.llm7.io/v1/chat/completions with the default selector. No GitHub token or provider API key is sent with that request.

LLM7.io currently documents anonymous limits of 60 requests per hour and 500,000 input-plus-output tokens per rolling 24 hours. Anonymous usage data may be processed for analysis and model improvement. The default route can choose a different underlying model between requests and has no service-level, availability, or reproducibility guarantee. When a quota, provider, response-validation, or routing failure occurs, the bot posts the complete deterministic report with a short fallback notice. See the official LLM7.io service information, anonymous limits, and model selector documentation.

The bot imports its trusted source tree directly and installs only a version- and hash-pinned PyYAML wheel on a GitHub-hosted Linux runner with Python 3.13. It does not resolve build dependencies during an issue-triggered job.

Security controls

  • Target URLs cannot select a host, port, credential, branch, ref, path, query, or fragment.
  • Only public repositories are accepted.
  • Only regular .yml, .yaml, and .md files directly under .github/workflows are fetched.
  • The target default branch is resolved once and all content calls use that full commit SHA.
  • At most 64 workflow files, 256 KiB per file, and 2 MiB total are accepted.
  • Target repositories are not cloned; hooks, submodules, LFS filters, dependencies, and code are never executed.
  • YAML source size, nesting, nodes, aliases, and expanded graph traversal are bounded.
  • The anonymous provider request contains only normalized language, scanned_files, counts, and catalog-backed rules aggregates.
  • Repository identity, commit SHA, raw issue text, workflow source, finding messages, traces, paths, GitHub tokens, and provider keys are not sent to LLM7.io.
  • The model destination is fixed to api.llm7.io/v1/chat/completions, and the selector is fixed to default; the underlying routed model can vary.
  • The scan, model, and comment jobs retain separate least-privilege GitHub permissions.
  • Model output is treated as untrusted, parsed and schema-checked locally, length-limited, mention-neutralized, and never used as a command, URL, identifier, or API target.

Operational limits

The issue form is public, so GitHub abuse controls and LLM7.io's anonymous limits are the practical request-rate boundaries. Those limits and the routed models may change, and LLM7.io provides no service-level guarantee for this anonymous route. A high-volume production service would require an external queue and per-actor rate limiter. The current bot is intended for public demonstrations and bounded repository checks. The global concurrency group limits simultaneous work, but it is not a rate limiter: GitHub keeps only one pending run in a group, so sustained issue spam can replace a legitimate pending scan and delay service. AI inference remains separately protected by the trusted-author/ai-approved gate, and its failure never changes the deterministic result.