mcp_clean_skill produces zero TP findings.

cleanup() can be called after resolve and does not raise.

`baseline` writes a file; scanning with it suppresses those findings.

When SKILLSPECTOR_PROVIDER=claude_cli, is_llm_available asks the provider.

scan with local directory runs graph and prints report.

scan with a --baseline pointing at a missing file exits with code 2.

scan with --no-llm runs without requiring an LLM API key (uses fallback).

scan with nonexistent path exits with code 2.

scan with --output writes report to file.

--version prints version and exits 0.

LLM response without start_line falls back to coarse (file, rule_id) keying.

HIGH severity finding in code-example context at low conf: retained by severity floor.

LOW severity finding at 0.6 conf in code-example context: 0.6*0.5=0.3 < 0.4, dropped.

Multiple rules are triggered when permissions, triggers, and parameters all change.

50 instances created concurrently inside one context — all get response_schema=None. V1 bug: class-attribute toggling across threads caused i

Ollama and some models return confidence on 0-100 scale; must be normalized.

Ollama-style 0-100 scale must be normalized to 0-1.

Out-of-range confidence is clamped, not rejected, so a slightly off model value does not fail the whole structured-output parse.

Out-of-range confidence is clamped, not rejected, so a slightly off model value does not fail the whole structured-output parse.

Values > 100 (e.g. 150) are divided then clamped.

Regression: a static finding with end_line=None must still match a confirmation whose end_line is populated (e.g. end_line == start_line, as s

#10: Tests n=2, not just n=1.

Verify module-level constants are in expected ranges.

Regression: the chat model is constructed INSIDE the try, so a construction failure degrades (records ok=False, preserves findings) instead of cra

Guard is called during deepseek_compat() enter — must succeed.

CRITICAL finding at 0.35 confidence is retained (severity floor).

A CRITICAL finding confirmed by the LLM is still enriched as before. The floor does not interfere with the normal happy path: when the LLM

Critical/harmful content produces high-severity findings.

A CRITICAL static finding NOT confirmed by the LLM must be kept. The finding must appear in the output with its original severity and

Cross-file dedup keeps the highest confidence finding.

A curl -k inside a markdown fenced code block should be filtered out.

Data URI in description → TP1 finding.

Same rule+line should not produce duplicate findings.

A degraded scan that is already DO_NOT_INSTALL stays blocking (floor only lifts SAFE).

No findings would normally be SAFE; a degraded LLM stage forces CAUTION.