Skip to content

Give the agent web context

Use sources for official changelogs, specifications, API references or vendor advisories that explain what the repository should support. This works in YAML and in SDK configuration objects.

version: 2
agent:
  allow_paid: true
sources:
  - https://developers.openai.com/api/docs/changelog
official_hosts:
  - developers.openai.com
instructions: |
  Compare documented API changes with this repository's implementation.
  Fix one concrete compatibility gap and add offline contract tests.
  Cite the reference evidence. Return unchanged if no fix is needed.
validation:
  checks:
    - name: tests
      command: [npm, test]

Adapt the checks and editable files for your repository; this example only shows how references fit into configuration.

What happens on a run

  1. The engine fetches each configured page over HTTPS.
  2. It extracts readable text and adds a bounded excerpt to the agent's brief.
  3. The agent receives the task, permitted paths, repository source and references.
  4. Changed patches must pass your independent checks before they are marked changed.

The saved brief.md artifact contains the supplied excerpts. Results can cite source URLs, but source citations alone do not prove a fix is correct.

Source boundaries

Sources must use allowed hosts. When official_hosts is omitted, it is derived from the source URLs; set it explicitly when you want a reviewed list. Redirects to other hosts must also pass the allowlist. Embedded credentials and non-HTTPS source URLs are rejected.

Use public, text-readable pages; there is no authenticated browser session or JavaScript rendering. Excerpts are bounded and may be truncated. The built-in Groq/Google profiles start with shorter excerpts to fit their context budgets. Failed or unavailable references stop preparation rather than silently running without the requested evidence.

The engine fetches current content on each run. It does not maintain a crawler, a search index, or a historical page-diff database. If a task needs more context, choose a more specific source or supply a concise task brief.

Web pages are evidence

A reference page, scanner finding or repository comment can contain misleading instructions. Treat it as evidence to compare with source and tests. Keep the task, editable boundaries and check commands in trusted repository configuration or your own script. Do not copy webhook-supplied URLs into the source allowlist.

For security fixes, include vendor advisories you trust in the configuration, and supply the individual finding's reproduction details as task evidence. See the vulnerability template.