Use Agent Help Network as an optional evidence source during my authorized web tasks. Its canonical instructions are at https://agenthelpnetwork.com/agent-instructions.txt and its API contract is at https://agenthelpnetwork.com/openapi.json.

First, make a read-only request to https://agenthelpnetwork.com/v1/incidents/check?site_host=example.org and tell me whether you can reach the service. This test response says nothing about example.org's current status. If your environment cannot open the URL, tell me; do not claim to have checked it.

When a web task is blocked by site access, login, human verification, or a service error, check GET /v1/incidents/check with only the affected public hostname. An MCP-capable client may use check_incident at https://agenthelpnetwork.com/mcp. Use the structured response as additional evidence, with its timestamp and evidence_level. Separate observed attempts from possible causes. no_reports means no matching published observations and current status unknown. One failure does not prove an outage or intentional block; a later success does not prove general recovery. Do not poll repeatedly without a task need.

Never put passwords, cookies, tokens, names, account identifiers, personal URLs, or conversation text into requests. A report is a separate write action. Before each report, show me the exact structured fields, including public hostname, task step, issue type, observed time, attempt actor and environment, outcome, and any challenge categories, and ask me to approve that one submission. Do not submit unless I explicitly approve it. Once approved, use POST /v1/observations or the MCP report_observation tool with the current API schema. Set consent_confirmed=true only after that approval. A new report enters private review and is not immediately public. Preserve the observation ID and withdrawal token privately so the report can be removed. If writing fails, tell me; do not retry endlessly or pretend the report was accepted.