Configuration

/tt setup creates .trigger-tree/config.sh. Every key resolves in three layers — bundled plugin default, then the user-wide ~/.trigger-tree/config.sh (location overridable via TT_USER_CONFIG), then the project file — and the project always wins. The user layer exists so one person can set a default such as TT_LOG_PROMPTS='off' for every repository before any project has run setup; tt doctor reports the effective prompt mode and which layer selected it.

Variable	Default	Meaning
TT_WATCH_REGEX	docs/agents/skills/briefs plus root CLAUDE.md, AGENTS.md, and GEMINI.md	Documentation reads to count
TT_SCOPE_IGNORE	empty	comma-separated globs acknowledging intentionally unwatched markdown; leaves gate findings, SARIF, and the watch-scope denominator
TT_SCAN_REGEX	documentation folders	Search targets to count
TT_ALWAYS_LOADED_REGEX	CLAUDE.md, AGENTS.md, GEMINI.md, guidance, and skills	Context excluded from cold-path review
TT_CRITICAL_GLOB	empty	Comma-separated rare-but-critical paths
TT_LOG_PROMPTS	hash	truncate, hash, or off for future prompt text/fingerprints; setup recommends truncate per project
TT_LOG_TOPICS	off	on stores at most eight matched labels from a repository-bounded vocabulary; no free prompt text
TT_LOG_COMMANDS	off	classified stores matched manifest pattern IDs only; full also stores a bounded command line; output is never stored
/tt setup writes a snapshot of the bundled defaults into the project, so improvements to
those defaults never reach an existing install on their own. tt doctor reports it when
your TT_ALWAYS_LOADED_REGEX predates the current version and therefore misses a surface
the client injects, because such a file is otherwise counted as untouched documentation
and lowers the health score. Copy the new pattern from the plugin's scripts/tt-config.sh.

| TT_LOG_AGENTS | off | on stores the persona name of a launched subagent; the task prompt and description are never read | | TT_EDIT_REGEX | matches nothing | Regex selecting in-project edit paths for adherence; events contain path and tool only | | TT_ROTATE_BYTES | 5 MB | History rotation threshold | | TT_EXPERIMENTAL_OUTCOMES | off | Local correlational committed/abandoned view |

For a new project, interactive setup asks for the prompt mode and recommends truncate; Enter accepts it. Setup also writes the optional adherence capture choices explicitly: topics, classified commands, and a conservative project-relative edit scope. Existing installs that have no new keys remain off and never start recording more silently. Piped, hook, and CI runs cannot block on the question and use privacy-preserving defaults. Existing project choices are preserved unless their explicit options are passed. Setup also reports watch coverage and can propose a regex, but never applies it without tt-setup.py --apply-watch-suggestion. /tt doctor fails on zero watch matches, warns on very low coverage, and identifies manifest probes that cannot fire under current capture.

TT_LOG_TOPICS is independent of TT_LOG_PROMPTS: topic matching behaves identically in truncate, hash, and off modes. The vocabulary comes only from router labels and manifest topics, matching case-insensitive whole words, capped at eight. Topic events can include a manifest fingerprint so evidence from a different vocabulary is not treated as negative evidence.

TT_LOG_COMMANDS=classified records an array such as "matched":["run-checks"], never the original command. full is a separate, explicit privacy choice and still never records stdout or stderr. TT_EDIT_REGEX can widen measurement beyond documentation; use the narrowest project-relative paths needed by command_after_edit and path_avoided probes. External and escaping paths are rejected, and edit content and diffs are never recorded.

All three root instruction conventions are watchable for consistent event handling, but they are classified as injected/always loaded and excluded from untouched review. index.md remains a normal folder router; it has no Gemini-specific classification.

External local tools can ingest a validated event with python3 <plugin>/scripts/tt-log.py ingest '{"t":"read","path":"docs/design/index.md"}'. Invalid events are dropped silently.

Publishing the docs-health badge

Run /tt badge (or python3 scripts/tt-stats.py --badge) to write a shields.io endpoint at .trigger-tree/badge.json. Because telemetry stays local and is gitignored, a normal CI checkout cannot calculate this grade. From the measured development checkout, run make badge-publish to update only docs-health.json on an existing badges branch while preserving its other endpoints.

[![docs health](https://img.shields.io/endpoint?url=https%3A%2F%2Fraw.githubusercontent.com%2FOWNER%2FREPO%2Fbadges%2Fdocs-health.json)](docs/heat-model.md)

CI independently updates coverage.json on the same branch and preserves the locally published docs-health value. Before measurement is mature, the public badge deliberately says measuring…. Publishing is explicit because it pushes local aggregate evidence to the repository; file paths and event history are never included in the badge payload.