Check codex is authenticated via ``codex login status`` (no inference).

Read *stream* into *buf* up to *cap* bytes, then stop reading. Calls *on_overflow* once if the cap is reached so the caller can react (e.g. k

Clear caches before and after every test. Depends on the conftest autouse ``mock_resolve_context_length`` fixture so it is active for the res

Gemini availability probe. The Gemini CLI (0.46.0) has no cheap non-interactive auth-status command, so we treat binary-on-PATH as available

Return a mock ChatOpenAI that supports with_structured_output.

Return a mock ChatOpenAI that supports with_structured_output.

Return a mock chat model that supports with_structured_output.

Run while main thread is inside deepseek_compat().

Extract assistant text from ``gemini -o json`` output. ``gemini -o json`` returns a JSON object with a ``response`` key (alongside ``session_

LLMs sometimes return nested objects as JSON strings.

LLMs sometimes return the findings array as a JSON string.

Set response_schema=None on the instance dict BEFORE original init. Relies on Python MRO guarantee: instance.__dict__ is always checked befor

Parse raw LLM text into Findings via manual JSON + Pydantic.

Parse raw LLM text into meta-analyzer dicts via manual JSON + Pydantic.

Scan a single skill through the full pipeline. Returns ------- (entry, error_message_or_None, relative_name)

Total active requests across all keys.

Analyze content for prompt injection patterns (P1–P4).

Analyze content for data exfiltration patterns (E1–E5).

Analyze content for server-side request forgery patterns (SSRF1–SSRF3).

Analyze content for excessive agency patterns (EA1–EA4).

Analyze content for rogue agent patterns (RA1–RA2).

Analyze content for anti-refusal statements (AR1-AR3).

Analyze content for output handling patterns (OH1–OH3).

Analyze content for tool misuse patterns (TM1–TM3).

Analyze content for privilege escalation patterns (PE1–PE5).

Analyze content for memory poisoning patterns (MP1–MP3).

Analyze content for supply chain patterns (SC1–SC3).

Analyze content for agent snooping patterns (AS1–AS3).

Analyze content for harmful content patterns (P5).

Analyze content for system prompt leakage patterns (P6–P8).

No side effects — returns the archive path string only.

``True`` when this key can accept at least one more caller.

Create a backup copy of the config file.

Minimal valid SkillspectorState for semantic_security_discovery.

Generate a baseline file that suppresses every finding in the current scan. Run this once to accept all existing findings, then commit the f

Walk directory and return a list of file metadata dicts (name, size, path). Read-only: no files are created, modified, or deleted.

Remove temporary and cache files without asking.

Invoke the CLI via the hardened runner and return the assistant text. The prompt is passed through unchanged (parity with the HTTP path).

Create ``ChatAnthropic`` with custom transport for the proxy.

Construct a ``ChatBedrockConverse`` bound to a Bedrock client. Returns ``None`` if no AWS credentials resolve in the boto3 credential

Create ``ChatAnthropic`` using native Anthropic credentials.

Create ``ChatOpenAI`` for the build.nvidia.com endpoint.

Create ``ChatOpenAI`` using standard OpenAI environment variables.

Deploy a service to the given environment.

Format a markdown file using an external tool.