Module classifier 

State classifiers: determine agent state from virtual screen content.

Each agent type (Claude, Codex, Kiro, Generic) has different prompt patterns, spinner indicators, and context exhaustion messages.

Structs

Classification

Enums

AgentType

Agent type selector for the shim classifier.

NarrationLineKind

How a single output line should be treated by narration enforcement.

ScreenVerdict

What the classifier thinks the agent is doing.

Constants

MIN_CLASSIFIER_CONFIDENCE

Functions

classify

Classify screen content based on agent type.

classify_narration_line

Classify a single output line for narration-loop detection.

classify_with_confidence

Classify screen content and return a confidence score for the match.

detect_claude_dialog

Legacy alias for backward compatibility in tests.

detect_meta_conversation

Detect meta-conversation patterns where the agent keeps planning or narrating without moving to concrete execution.

detect_narration_pattern

Detect whether the current screen content looks like narration instead of concrete execution. This is intentionally screen-level so daemon health checks can count consecutive narration polls.

detect_startup_dialog

Detect known startup dialogs that should be auto-dismissed by the shim. Works for Claude, Codex, and other agents that show trust prompts.