Privacy and local data

• Runtime code makes no network calls and uses only Python’s standard library.
• Telemetry stays in the project’s gitignored .trigger-tree/ directory.
• Read contents, search patterns/output, MCP responses, edit contents/diffs, and command output are not stored.
• Before a project's setup, the fallback is hash: a short stable SHA-1 fingerprint, never prompt text. Plugin hooks are user-wide, so this protects repositories that have not chosen a mode yet. A user-wide default in ~/.trigger-tree/config.sh (or the file TT_USER_CONFIG points to) can tighten this to off for every repository before any setup; a project's own choice always wins. /tt setup asks per project: truncate (recognizable 200-character previews, stored locally and gitignored), hash, or off (marker only). Changes affect future prompts only.
• Deletion belongs to the user. Uninstall removes wiring but intentionally preserves telemetry.

The experimental outcome view observes local HEAD changes and test-command results. It is correlational and off by default. See the complete privacy policy and security policy.

Instruction adherence adds three explicit, independently controlled capture surfaces. All default to off for an existing install with no configuration:

• TT_LOG_TOPICS=on: up to eight whole-word topic labels drawn only from the repository's router/manifest vocabulary. No prompt-derived free text reaches disk, regardless of TT_LOG_PROMPTS.
• TT_LOG_COMMANDS=classified: only stable IDs of manifest patterns that matched.
• TT_LOG_AGENTS=on: the persona name of a launched subagent, nothing else. The launch payload also carries the task prompt and a description; neither is read. full additionally stores a bounded command line; neither mode stores output.
• TT_EDIT_REGEX: matching edits store only normalized in-project path, tool name, and event metadata. External paths, content, arguments unrelated to paths, and diffs are not retained.

The committed .trigger-tree/directives.json is configuration, not telemetry. It contains instruction paths and hashes, source references, optional directive text, and user-confirmed probes. Local analysis reads instruction content to verify hashes and estimate cost, but does not copy that content into event history or send it anywhere.