Detectors

Detectors

Sentinel ships with 14 built-in detectors organized by tier. Each detector produces evidence-backed findings with severity, confidence, and file location.

Detector tiers

Tier	Description	LLM needed?
DETERMINISTIC	Rule-based, repeatable, zero false negatives from judgment	No
HEURISTIC	Statistical/AST analysis, high confidence but uses thresholds	No
LLM_ASSISTED	Uses an LLM for semantic comparison or judgment	Yes (BASIC tier minimum)

All detectors

Detector	Tier	Languages	External Tool	LLM	Page
Detector: Todo Scanner	DETERMINISTIC	All	git blame	No	Detector: Todo Scanner
Detector: Dep Audit	DETERMINISTIC	Python	pip-audit	No	Detector: Dep Audit
Detector: Lint Runner	DETERMINISTIC	Python	ruff	No	Detector: Lint Runner
Detector: ESLint Runner	DETERMINISTIC	JS/TS	Biome or ESLint	No	Detector: ESLint Runner
Detector: Go Linter	DETERMINISTIC	Go	golangci-lint	No	Detector: Go Linter
Detector: Rust Clippy	DETERMINISTIC	Rust	cargo clippy	No	Detector: Rust Clippy
Detector: Stale Env	DETERMINISTIC	Python, JS/TS	—	No	Detector: Stale Env
Detector: Unused Deps	DETERMINISTIC	Python, JS/TS	—	No	Detector: Unused Deps
Detector: Complexity	HEURISTIC	Python	—	No	Detector: Complexity
Detector: Git Hotspots	HEURISTIC	All	git log	No	Detector: Git Hotspots
Detector: Dead Code	HEURISTIC	Python, JS/TS	—	No	Detector: Dead Code
Detector: Docs Drift	LLM_ASSISTED	All (docs)	—	Optional	Detector: Docs Drift
Detector: Semantic Drift	LLM_ASSISTED	All (docs)	—	Required (BASIC)	Detector: Semantic Drift
Detector: Test Coherence	LLM_ASSISTED	Python	—	Required (BASIC)	Detector: Test Coherence

Note on LLM_ASSISTED tier: Docs-drift is tiered as LLM_ASSISTED but degrades gracefully — its deterministic sub-checks (stale references, dependency drift) run without any LLM. Only the doc-code semantic comparison requires a model. Semantic-drift and test-coherence produce no findings without an LLM provider.

Enabling/disabling detectors

Via config

[sentinel]
# Run only these detectors:
enabled_detectors = ["todo-scanner", "complexity", "dep-audit"]

# Or skip specific detectors:
disabled_detectors = ["lint-runner", "dead-code"]

enabled_detectors and disabled_detectors are mutually exclusive.

Via CLI

sentinel scan /repo --detectors todo-scanner,complexity,dep-audit
sentinel scan /repo --skip-detectors lint-runner,dead-code

Via profiles

sentinel init /repo --profile minimal   # heuristic-only, no LLM
sentinel init /repo --profile standard   # all detectors, basic LLM
sentinel init /repo --profile full       # all detectors, enhanced analysis

How detectors work

1. The runner discovers all registered detectors (built-in + entry-point plugins)
2. Filters by enabled/disabled lists
3. Each detector receives a DetectorContext with repo root, scope, changed files, config, and optional model provider
4. Each detector returns a list of Finding objects with evidence
5. Findings are fingerprinted, deduplicated, and optionally judged by the LLM

See Writing Custom Detectors for building your own.