Skip to content

Privacy modes

A privacy mode decides what a model provider may see. It applies to every model call: summaries, suggestions, commit intent inference, Conventional Commit drafting and the MCP evidence tools.

Mode Local providers (ollama, mock, stub) Remote providers
metadata-only (default) structure only structure only
local-summary-then-ai structure and source content structure only
trusted-provider-patch structure and source content structure and source content

Structure means the change summary, PR and commit metadata, file paths and statuses, deterministic diff facts, and IntentumDiff's semantic diff with identifier names: which definitions changed, were added, removed, renamed, moved or refactored, and which edits were formatting only. Source content adds IntentumDiff's change descriptions, which can quote a changed statement before and after (bounded to 160 characters per change), and the source text of added or deleted definitions (bounded to 120 characters). With the legacy git engine, source content means raw patch excerpts.

Under local-summary-then-ai, source content goes only to local providers, and only if every provider in the fallback chain is local too. trusted-provider-patch sends source content to any provider, local or remote: choose it only for providers your team trusts with your code. --allow-remote-patches and the verlato.allowRemotePatches setting are explicit opt-ins that select it (for scans and drafts, so does VERLATO_ALLOW_REMOTE_PATCHES=1).

Setting the mode

In order of precedence: --allow-remote-patches (selects trusted-provider-patch), then --privacy <mode>, then VERLATO_MODEL_PRIVACY (the verlato.modelPrivacy setting), then model_privacy in verlato.json. Unknown values fall back to metadata-only.

Seeing what is sent

Every model call records what it sent:

verlato draft prompt --from-latest-tag --id ci_715c52ad2c02

returns the exact prompt and evidence a suggestion would send, without calling a model, plus its privacy record:

"privacy": {
  "mode": "metadata-only",
  "provider": "ollama",
  "raw_patch_chars_available": 659,
  "raw_patch_chars_sent": 0,
  "will_send": ["change summary", "PR metadata", "commit metadata", "file paths and statuses", "deterministic diff facts"],
  "will_not_send": ["raw patches", "patch excerpts", "representative hunks"]
}

Model answers carry an audit record too (ai_request), with the SHA-256 hashes of the request, prompt and evidence, the prompt and evidence sizes, and what was omitted and why. Applied suggestions keep this provenance with the item.

Suggestion strategies and privacy

The batched (paired, full-diff-json) and tool-retrieval suggestion strategies exist to show models more of each change, so they only run when the mode allows source content for the provider. Otherwise Verlato quietly uses the single strategy.

MCP clients

MCP clients are treated as remote models: IntentumDiff's change descriptions are only included in MCP answers under trusted-provider-patch. See The MCP server.