agbrowse docs

Guide

Fail closed when completed answers lack required source coverage.

Source audit checks provider answer text for inline source coverage and can block completed answers when a supervising workflow requires cited claims.

Query with source audit

agbrowse web-ai query \
  --vendor chatgpt \
  --inline-only \
  --require-source-audit \
  --source-audit-ratio 1 \
  --prompt "Find the latest official installation docs and cite them."

Scoped audit

Narrow what counts as a valid source and when. Use --source-audit-scope and --source-audit-date for absence claims such as "no official response was found".

agbrowse web-ai query \
  --vendor grok \
  --model expert \
  --inline-only \
  --require-source-audit \
  --source-audit-scope "official product docs and release notes" \
  --source-audit-date "2026-05-05" \
  --prompt "Summarize the latest official product changes with sources."

Use --source-audit-ratio <0..1> only when partial sourcing is deliberate; the default requires every detected claim to carry an inline source.

Claim audit

agbrowse web-ai claim-audit scans repository docs for forbidden hosted/cloud/stealth claims. It uses the same claim-audit module as the release gate gate:no-cloud-claims, ensuring that public documentation never makes claims the codebase cannot back.

agbrowse web-ai claim-audit --json

Integration with release gates

The gate:no-cloud-claims release gate runs claim-audit automatically during npm run gate:all. Any forbidden claim found in README, docs, or structure files causes the gate to fail, blocking release. This keeps public-facing text aligned with the capability truth table in structure/CAPABILITY_TRUTH_TABLE.md.

npm run gate:no-cloud-claims   # run the claim gate standalone
npm run gate:all               # includes claim audit with all other gates

Live status report

The web-ai subsystem emits typed standalone live-status report structs during long-running operations. These structured reports carry provider state, progress signals, and timing data that downstream tools can parse without scraping CLI output.